Corrosion monitoring method and system for carbon dioxide capture equipment

Abstract

The invention provides a corrosion monitoring method and system for carbon dioxide capture equipment, and relates to the technical field of corrosion monitoring, and the method comprises the steps: obtaining the real-time operation data, operation time and corrosion degree before operation of the current carbon dioxide capture equipment; obtaining the final corrosion degree of equipment of the same type and training data including full-life-cycle historical operation data and corresponding corrosion rates; establishing a neural network model by using a machine learning algorithm, and training the model by using the training data; inputting real-time operation data into the trained model, and outputting a predicted corrosion rate; the remaining life of the device is calculated based on the final corrosion degree, the pre-run corrosion degree, the run time, and the predicted corrosion rate. The corrosion state can be monitored under the non-stop condition, the problem that a traditional method lags in response to a complex corrosion mechanism of a chemical absorption system is solved, and the monitoring precision and reliability of long-period operation equipment are remarkably improved.

Description

Technical Field

[0001] The present application relates to the technical field of corrosion monitoring, and in particular, to a method and system for monitoring the corrosion of carbon dioxide capture equipment. Background Art

[0002] In response to the urgent needs of global climate change and environmental protection, carbon dioxide capture and storage (CCS) technology has become one of the key means to reduce greenhouse gas emissions. The chemical absorption method is one of the most widely used and mature carbon capture technologies in industrial applications. In the chemical absorption method, organic amines have the advantages of fast absorption rate and large absorption capacity, and are the preferred absorbents. However, when carbon dioxide dissolves in water or organic amine solution, it will react to form carbamate (pH drops from 10 - 12 to 6 - 8), causing the dissolution of the Fe3O4 passive film on carbon steel equipment (Fe3O4 + 8H + →3Fe 2+ + 4H2O); the heat - stable salts (HSS) generated by the degradation of amine solution at high temperature (110 - 130°C) in the regeneration tower result in a local pitting corrosion rate of 2 - 5 mm / year, and the temperature alternation (40°C - 120°C) and CO2 loading rate fluctuation (0.2 - 0.8) caused by the rich - lean liquid circulation will further induce stress corrosion cracking.

[0003] Traditional corrosion monitoring methods mainly perform regular sampling and analysis of the corrosion degree through ultrasonic monitoring, and then calculate the real - time corrosion rate based on the sampling time difference (or monitor the corrosion rate in real - time through electrochemical monitoring methods); then, based on the real - time corrosion rate and the remaining corrosion degree of the equipment, the remaining life of the equipment is calculated.

[0004] When using the electrochemical monitoring method to monitor the corrosion rate in real - time, if the corrosion rate is small or the corrosion rate change rate is small, the sensor cannot perceive the change in the corrosion rate. Over time, the monitoring error of the corrosion degree will have a drift phenomenon. To overcome the drift phenomenon, ultrasonic monitoring can be supplemented to monitor the corrosion degree, but ultrasonic monitoring requires shutdown inspection and is not suitable for the inspection of carbon dioxide capture equipment with a long working cycle. Therefore, for carbon dioxide capture equipment with a long working cycle, it is difficult to ensure the monitoring accuracy by only using the electrochemical monitoring method to monitor the corrosion rate and the remaining life. Summary of the Invention

[0005] In order to improve the accuracy of the corrosion state monitoring results of carbon dioxide capture equipment with a long working cycle, the present application provides a method and system for monitoring the corrosion of carbon dioxide capture equipment.

[0006] In the first aspect, the present application provides a method for monitoring the corrosion of carbon dioxide capture equipment, adopting the following technical solution: A method for corrosion monitoring of a carbon dioxide capture device, comprising: Real-time data acquisition: Obtain the real-time operation data, operation time, and corrosion degree before operation of the current carbon dioxide capture device; Historical data acquisition: Obtain the final corrosion degree and training data of the same type of carbon dioxide capture device. The training data includes: historical operation data of the same type of carbon dioxide capture device throughout its life cycle, and the corrosion rate of the same type of carbon dioxide capture device under each historical operation data; Modeling and training: Use a machine learning algorithm to establish a neural network model, and train the neural network model with the training data to obtain a trained neural network model; Predict the corrosion rate: Input the real-time operation data into the trained neural network model, and output the predicted corrosion rate; Calculate the remaining life: Calculate the remaining life based on the final corrosion degree, the corrosion degree before operation, the operation time, and the predicted corrosion rate.

[0007] This application first obtains the real-time operation data, operation time, and corrosion degree before operation of the current carbon dioxide capture device. Among them, the real-time operation data reflects the working state of the current carbon dioxide capture device, and the corrosion degree before operation is combined with the operation time to calculate the current corrosion degree of the device. Subsequently, this application obtains the final corrosion degree and training data of the same type of carbon dioxide capture device. The final corrosion degree of the same type of carbon dioxide capture device is used as the end point of the life, which is used to calculate the remaining life in combination with the current corrosion degree. Subsequently, this application uses a machine learning algorithm to establish a neural network model and trains the neural network model with the training data. During the training process, the neural network model can automatically learn the potential laws and features in the data, thereby improving the prediction performance. Subsequently, this application inputs the real-time operation data into the trained neural network model, and the trained neural network model will output the predicted corrosion rate. This application can calculate the current corrosion state through the operation time and the corrosion degree before operation, and is applicable to carbon dioxide capture devices with long-cycle work that require accurate calculation of the remaining life.

[0008] Optionally, the historical operation data includes m types of operation indicators. After performing the step of historical data acquisition and before performing the step of modeling and training, it further includes: Clustering: Use the K-means clustering algorithm to cluster the historical operation data under the i-th type of operation indicator to obtain clusters, where the value of K is 2; Determine the threshold: Maximize the between-class variance of the two clusters by the large law method, and record the historical operation data corresponding to the maximum between-class variance as the alarm threshold for the i-th type of operation indicator; Threshold judgment: sequentially determine whether each piece of real-time operation data is greater than the corresponding alarm threshold. If so, send a signal indicating that the data is unqualified; if not, perform the steps of modeling and training.

[0009] This application uses the K-means clustering algorithm to cluster the historical operation data under each operation index (K is taken as 2). The clustering aims to gather the normal operation data and abnormal operation data in different clusters respectively. In this way, the complex data set can be simplified into two representative clusters, which helps to discover the potential patterns and distribution characteristics in the data. Subsequently, this application determines the alarm threshold by maximizing the between-class variance of the two clusters using Otsu's method. Otsu's method can automatically find an optimal segmentation point in the data to maximize the difference between the two clusters, thereby determining a reasonable alarm threshold. This method of determining the threshold is more scientific and accurate than subjectively setting the threshold and can better reflect the actual situation of the data. This application also sets alarm thresholds for each operation index, fully considering the characteristics and differences of different indexes. Different operation indexes may have different data distributions and variation ranges. Determining personalized alarm thresholds for each index can improve the pertinence and effectiveness of the thresholds. Subsequently, this application sequentially determines whether the real-time operation data is greater than the corresponding alarm threshold. Once it is found that the data is unqualified, a signal indicating that the data is unqualified is immediately sent, so as to timely discover the abnormal situation in the equipment operation process.

[0010] Optionally, after performing the step of determining the threshold and before performing the step of threshold judgment, it further includes: Calculating the average value: calculate the average value of the historical operation data under each operation index respectively; Data analysis: adopt the method similar to PDP, and sequentially input the historical operation data under the i-th operation index and the average values corresponding to the remaining operation indexes into the neural network model to obtain the change amount of the prediction result of the i-th operation index. The calculation model of the change amount of the prediction result of the i-th operation index is as follows: ; where is the output probability when inputting the j-th historical operation data in the i-th operation index; is the output probability of the k-th remaining operation index when inputting the j-th historical operation data in the i-th operation index; n is the number of historical operation data in the i-th operation index; m is the number of operation indexes; Obtaining the change trend: obtain all the change amounts, and set the priorities of the operation indexes according to the order of the change amounts from large to small; In the step of threshold judgment, compare the real-time operation data with the corresponding alarm threshold according to the priority.

[0011] This application calculates the average value of historical operation data under each operation index. Then, using the method similar to PDP (Partial Dependence Plot), the historical operation data under the i-th operation index and the average values corresponding to the remaining operation indexes are input into the neural network model in sequence to obtain the change amount of the prediction result of the i-th operation index, so as to quantify the independent influence of each operation index on the prediction result and help identify which operation indexes are more critical for predicting the corrosion rate or remaining life of the equipment. By calculating the change amount, it is possible to more intuitively understand how the change of each operation index leads to the change of the prediction result. Subsequently, this application obtains all the change amounts and sets the priorities of the operation indexes according to the order from large to small of the change amounts, so that when making threshold judgments, it can be more targeted and give priority to considering the operation indexes that have a greater impact on the prediction result. For those operation indexes with larger change amounts and significant influence on the prediction result, a higher priority is given for threshold judgment, which can more effectively discover potential problems and improve the efficiency and accuracy of equipment monitoring. In the step of threshold judgment, the real-time operation data is compared with the corresponding alarm thresholds according to the priorities. This way of judging by priorities can more quickly detect abnormal situations of key operation indexes. Once it is found that the data of high-priority indexes exceeds the alarm threshold, measures can be taken in time to improve the response speed of equipment maintenance.

[0012] Optionally, the method further includes: Index screening: Perform Pareto analysis on the change amount of each operation index to obtain the key operation indexes, and update the key operation indexes as the operation indexes.

[0013] This application screens the change amount of each operation index through Pareto analysis to identify the key operation indexes that have the most significant influence on the prediction result. Among many operation indexes, usually only a few indexes play a decisive role in predicting the corrosion rate or remaining life of the equipment. Pareto analysis can quickly find these key indexes, thereby improving the efficiency of data analysis. This application updates the screened key operation indexes as the operation indexes for subsequent analysis, reducing the data dimension to be processed. In subsequent steps such as obtaining the change trend and making threshold judgments, only the key indexes need to be concerned, reducing the computational complexity and making the whole process more concise and efficient.

[0014] Optionally, after performing the step of predicting the corrosion rate and before performing the step of calculating the remaining life, it further includes: Environment acquisition: Acquire solvent environment data and corrosion current data with the same timestamp as the historical operation data; acquire the solvent environment data at the current moment; Modeling: Establish a random forest model, and use the solvent environment data and corrosion current data to train the random forest model to obtain the trained random forest model; Predict the corrosion current: Input the solvent environment data at the current moment into the trained random forest model to output the predicted corrosion current data; Calculate the corrosion rate: Calculate the theoretical corrosion rate through Faraday's law, and calculate the corrosion rate difference based on the theoretical corrosion rate and the predicted corrosion rate; Corrosion rate judgment: Judge whether the predicted corrosion rate difference is greater than the preset corrosion rate threshold. If so, send an alarm signal; if not, execute the step of calculating the remaining life.

[0015] This application obtains the solvent environment data and corrosion current data with the same timestamp as the historical operation data and uses these data in the modeling process, which can more comprehensively consider the factors affecting the corrosion process. By introducing the solvent environment data, the random forest model can learn the variation law of the corrosion current under different environmental conditions, so as to better adapt to the complex and changeable environment in actual operation. Subsequently, this application obtains the solvent environment data at the current moment and inputs it into the trained random forest model to predict the corrosion current data in real time. Then, the theoretical corrosion rate is calculated through Faraday's law, and the corrosion rate difference is calculated based on the theoretical corrosion rate and the predicted corrosion rate. This method combines theoretical calculation with actual prediction, which can more accurately reflect the actual situation of equipment corrosion. The theoretical corrosion rate provides a benchmark for prediction, while the corrosion rate difference reflects the deviation between the actual corrosion situation and the theoretical situation. Then, it is judged whether the predicted corrosion rate difference is greater than the preset corrosion rate threshold. When the corrosion rate difference exceeds the preset corrosion rate threshold, an alarm signal is sent in time to remind the management personnel that the equipment may have abnormal conditions and further inspection and maintenance are needed. Only when the corrosion rate difference does not exceed the threshold, this application will execute the step of calculating the remaining life to accurately evaluate the remaining service life of the carbon dioxide capture equipment under normal corrosion conditions.

[0016] Optionally, the random forest model includes p decision trees. After performing the step of environment acquisition and before performing the step of modeling, it further includes: Construct a data set: Integrate all the solvent environment data into a solvent environment data set, construct p - 1 sub - data sets, and randomly take a solvent environment data and put it into any one of the sub - data sets; Iteration: Put the selected solvent environment data back into the solvent environment data set and re - execute the step of constructing the data set until the number of samples in all sub - data sets is greater than the number of samples in the solvent environment data set, and update all sub - data sets to the solvent environment data set; In the step of modeling, the p - th solvent environment data set is used to train the p - th decision tree in the random forest model.

[0017] When constructing the data set in this application, all solvent environment data are integrated into a solvent environment data set, and then p - 1 sub - data sets are constructed, and the solvent environment data are randomly placed into the sub - data sets. By means of random assignment, this application increases the diversity of data, so that each sub - data set has a different sample combination. By increasing data diversity, each decision tree in the random forest model can access a wider range of data samples during training, reducing the risk of overfitting of a single decision tree to specific data. Different decision trees are trained based on different sub - data sets, learning different data features and rules, thus improving the generalization ability of the entire random forest model and enabling it to make more accurate predictions when facing new solvent environment data. During the iterative process, this application puts the selected solvent environment data back into the solvent environment data set and repeats the steps of constructing the data set until the number of samples in all sub - data sets is greater than the number of samples in the solvent environment data set. In this way, each solvent environment data has the opportunity to be assigned to different sub - data sets multiple times and thus be utilized by multiple decision trees. This application fully exploits the information in the data, improves the utilization rate of the data, and helps the model learn more comprehensive data features. Since each decision tree is trained based on a different sub - data set, the random forest model's dependence on the data is reduced. Even if there are some outliers or noises in a certain sub - data set, it will not have too much impact on the entire random forest model. This stability enables the random forest model to maintain good performance under different data distributions and working conditions.

[0018] Optionally, the step of constructing the data set further includes: dividing the solvent environment data set into q layers, randomly generating a random number x from 1 to q, and taking any solvent environment data from the x - th layer and putting it into any sub - data set.

[0019] This application divides the solvent environment dataset into q layers, and then extracts data from the divided q layers into a sub-dataset, so that data from different layers have the opportunity to be extracted into the sub-dataset, enabling the sub-dataset to more comprehensively cover various features in the solvent environment dataset. Subsequently, this application randomly selects a solvent environment data from the x layer and puts it into the sub-dataset, increasing the randomness of data extraction, which helps to break the possible correlations in the original dataset and makes the data in the sub-dataset more independent. When the decision tree is trained based on these independent sub-datasets, it can learn a wider range of data features and patterns, reduce the dependence on specific data patterns, thereby improving the generalization ability of the model and enabling it to make more accurate predictions when facing new and unseen solvent environment data. Randomly extracting data can prevent the model from overfitting to specific data samples during training. Since each sub-dataset is obtained by random extraction, the model needs to adapt to different data combinations rather than just memorizing certain specific samples, which helps to improve the robustness of the model. Different decision trees can learn different data features based on different sub-datasets, so that more abundant and accurate prediction results can be provided during model integration.

[0020] Optionally, the method further includes: Lifetime judgment: Judge whether the remaining lifetime is less than a preset lifetime threshold. If so, execute the warning step; if not, re-execute the step of real-time data collection. Warning: Send a warning signal.

[0021] By judging whether the remaining lifetime is less than the preset lifetime threshold, this application can timely send a warning signal when the remaining lifetime of the device is insufficient, enabling the device management personnel to understand the potential risks of the device in advance and take corresponding measures. The lifetime judgment and warning steps, together with steps such as real-time data collection and remaining lifetime assessment, form a closed-loop management system. By continuously collecting data, assessing the remaining lifetime, sending warnings, and handling problems, the device management process can be continuously optimized. The management personnel can summarize experience and lessons based on the warning information and the device operation situation, adjust the management strategies and methods, and improve the scientificity and effectiveness of device management.

[0022] In a second aspect, this application provides a corrosion monitoring system for a carbon dioxide capture device, adopting the following technical solution: A corrosion monitoring system for a carbon dioxide capture device, comprising: A monitoring module, including a real-time data collection unit and a historical data collection unit; The real-time data collection unit is used to obtain the real-time operation data, operation time, and corrosion degree before operation of the current carbon dioxide capture device. A historical data acquisition unit for obtaining the final corrosion degree and training data of the same type of carbon dioxide capture equipment, where the training data includes: historical operation data of the same type of carbon dioxide capture equipment throughout its life cycle, and the corrosion rate of the same type of carbon dioxide capture equipment under each historical operation data; An edge computing module, including a modeling and training unit, a predicted corrosion rate unit, and a remaining life calculation unit; The modeling and training unit is used to establish a neural network model using a machine learning algorithm, and train the neural network model with the training data to obtain a trained neural network model; The predicted corrosion rate unit is used to input real-time operation data into the trained neural network model and output a predicted corrosion rate; The remaining life calculation unit is used to calculate the remaining life based on the final corrosion degree, the corrosion degree before operation, the operation time, and the predicted corrosion rate.

[0023] Optionally, the system further includes: A memory, which stores a computer-readable storage medium and training data; A processor, which executes the steps of the method when calling the program code in the memory.

[0024] In summary, the present application includes at least one of the following beneficial technical effects: 1. The present application first obtains the real-time operation data, operation time, and corrosion degree before operation of the current carbon dioxide capture equipment. Among them, the real-time operation data reflects the working state of the current carbon dioxide capture equipment, and the corrosion degree before operation is combined with the operation time to calculate the current corrosion degree of the equipment. Subsequently, the present application obtains the final corrosion degree and training data of the same type of carbon dioxide capture equipment. The final corrosion degree of the same type of carbon dioxide capture equipment is used as the end point of the life to calculate the remaining life in combination with the current corrosion degree. Subsequently, the present application uses a machine learning algorithm to establish a neural network model and trains the neural network model with the training data. During the training process, the neural network model can automatically learn the potential laws and features in the data, thereby improving the prediction performance. Subsequently, the present application inputs the real-time operation data into the trained neural network model, and the trained neural network model will output a predicted corrosion rate. The present application can calculate the current corrosion state through the operation time and the corrosion degree before operation, and is applicable to carbon dioxide capture equipment with a long cycle of work that requires accurate calculation of the remaining life.

[0025] 2. This application uses the K-means clustering algorithm to cluster the historical operation data under each operation index (K is taken as 2). The clustering aims to gather the normal operation data and abnormal operation data in different clusters respectively. In this way, the complex data set can be simplified into two representative clusters, which helps to discover the potential patterns and distribution characteristics in the data. Subsequently, this application determines the alarm threshold by maximizing the between-class variance of the two clusters through Otsu's method. Otsu's method can automatically find an optimal segmentation point in the data to maximize the difference between the two clusters, thereby determining a reasonable alarm threshold. This method of threshold determination is more scientific and accurate than subjectively setting the threshold and can better reflect the actual situation of the data. This application also sets alarm thresholds for each operation index, fully considering the characteristics and differences of different indexes. Different operation indexes may have different data distributions and variation ranges. Determining personalized alarm thresholds for each index can improve the pertinence and effectiveness of the thresholds. Subsequently, this application successively determines whether the real-time operation data is greater than the corresponding alarm threshold. Once unqualified data is found, a data unqualified signal is immediately sent, so as to timely discover the abnormal situation in the equipment operation process. Description of the Drawings

[0026] Figure 1 is the flowchart of Embodiment 1 of this application; Figure 2 is the flowchart of Embodiment 2 of this application; Figure 3 is the flowchart of Embodiment 3 of this application. Detailed Description of the Embodiments

[0027] The following is combined with Figures 1 to 3 to further elaborate on this application in detail.

[0028] The carbon dioxide capture equipment mentioned in this application includes an absorption tower, a desorption tower, a heat exchanger, a pump, a compressor, a separator, and a control system.

[0029] Embodiment 1: This embodiment discloses a method for monitoring the corrosion of a carbon dioxide capture equipment. Refer to Figure 1, the method includes: S11 real-time data collection, S12 historical data collection, S13 modeling and training, S14 predicting the corrosion rate, and S15 calculating the remaining life. In this embodiment, first, the real-time operation data, operation time, and pre-operation corrosion degree of the current carbon dioxide capture device are obtained; the final corrosion degree of the same type of device and the training data including the historical operation data and corresponding corrosion rates throughout the life cycle are obtained; then, a neural network model is established using a machine learning algorithm, and the model is trained with the training data; then, the real-time operation data is input into the trained model to output the predicted corrosion rate; finally, based on the final corrosion degree, pre-operation corrosion degree, operation time, and predicted corrosion rate, the remaining life of the device is calculated. The process of this embodiment is as follows: S11 Real-time data collection, obtaining the real-time operation data, operation time, and pre-operation corrosion degree of the current carbon dioxide capture device.

[0030] The operation data reflects the current operation state of the carbon dioxide capture device, which covers the following aspects: Process parameters: including temperature, pressure, flow rate, temperature gradient ( ), O2 content; Device parameters: including device material parameters.

[0031] Record the duration (i.e., operation duration) experienced by the device from the start time of this operation to the current moment, and use the ultrasonic detection method to measure the corrosion degree of the current carbon dioxide capture device before this operation. The corrosion degree refers to the corrosion wall thickness, that is, the thickness of the carbon steel passivation film corrosion.

[0032] S12 Historical data collection, obtaining the final corrosion degree and training data of the same type of carbon dioxide capture device. The training data includes: the historical operation data of the same type of carbon dioxide capture device throughout the life cycle, and the corrosion rate of the same type of carbon dioxide capture device under each historical operation data.

[0033] Collect the corrosion degree data reached by the same type of carbon dioxide capture device at the end of the entire life cycle. These data can be from the retired or scrapped carbon dioxide capture devices. The ultrasonic detection is used to obtain the corrosion amount of each part; calculate the average value as the benchmark of the final corrosion degree.

[0034] The training data is the key to establishing an accurate model, which includes two main parts: The historical operation data of the same type of carbon dioxide capture device throughout the life cycle. The historical operation data records the operation parameters at each stage during the entire process from the commissioning to the retirement of the same type of carbon dioxide capture device. Similar to the real-time operation data, the historical operation data also includes the following various parameters: Process parameters: including temperature, pressure, flow rate, temperature gradient ( ), O2 content; Equipment parameters: including equipment material parameters.

[0035] However, the difference is that it covers the changes during the entire life cycle of the equipment and can reflect the characteristics and laws of the carbon dioxide capture equipment at different operation stages.

[0036] The corrosion rate of carbon dioxide capture equipment of the same type under each historical operation data. The corrosion rate refers to the change rate of the degree of equipment corrosion over time. By combining historical operation data and the corresponding corrosion rate, the correlation between the equipment operation state and the corrosion situation can be established.

[0037] S13 Modeling and training, using machine learning algorithms to establish a neural network model. The initial parameters of the neural network model are shown in Table 1.

[0038] Table 1 Initial parameters of the neural network model Network depth Convolution kernel Activation function Number of iterations Max pooling 2 64，5×5 ReLU 128 2×2 Input the collected training data into the neural network model. By continuously adjusting the parameters inside the model, the output result of the model can be made as close as possible to the actual corrosion rate.

[0039] During the training process, the stochastic gradient descent algorithm is used to minimize the prediction error of the model. During the training process, the training data will be divided into multiple batches, and a part of the data is used for training each time. The model parameters are updated through multiple iterations until the performance of the model reaches a satisfactory level. After training, the trained neural network model is obtained, and this model has learned the potential relationship between the equipment operation state and the corrosion rate.

[0040] S14 Predict the corrosion rate, input the real-time operation data into the trained neural network model. The trained neural network model will calculate and analyze the input real-time data according to the relationship between the carbon dioxide capture equipment operation state and the corrosion rate learned before, and finally output the predicted corrosion rate.

[0041] S15 Calculate the remaining life. According to the corrosion degree before operation and the final corrosion degree, the maximum allowable corrosion amount of the carbon dioxide capture equipment during the entire life cycle can be determined. Combining the operation time and the predicted corrosion rate, calculate how long it will take for the carbon dioxide capture equipment to reach the final corrosion degree at the current corrosion rate. This time is the remaining life of the equipment.

[0042] In this embodiment, real-time data collection is first performed to obtain the real-time operation data, operation time, and pre-operation corrosion degree of the current carbon dioxide capture device. Then, historical data collection is carried out to collect the final corrosion degree of the same type of device and the training data including the historical operation data and corresponding corrosion rates throughout the life cycle. Subsequently, a neural network model is established using a machine learning algorithm, and the model is trained with the training data. After that, the real-time operation data is input into the trained model to output the predicted corrosion rate. Finally, based on the final corrosion degree, pre-operation corrosion degree, operation time, and predicted corrosion rate, the remaining life of the device is calculated, so as to realize the prediction of the corrosion situation of the carbon dioxide capture device and the evaluation of the remaining life, providing decision-making support for device operation management.

[0043] In other embodiments, the method further includes: Life judgment, judging whether the remaining life is less than a preset life threshold. If so, the warning step is executed; if not, S11 real-time data collection is performed again.

[0044] Warning, sending out a warning signal.

[0045] Embodiment 2: Refer to Figure 2 , the difference between this embodiment and Embodiment 1 is that after performing S12 historical data collection and before performing S13 modeling and training, it further includes: S20 Index screening, performing a Pareto analysis on the change amount of each operation index. The Pareto analysis is an analysis method based on the "80 / 20 principle", that is, a few key factors often have a major impact on the result. In this step, the change amounts of all operation indexes are sorted to find out the key operation indexes that have a greater impact on the prediction result.

[0046] Through the Pareto analysis, it is possible to determine which operation indexes play a key role in the prediction result of the model, and these key operation indexes are used as the operation indexes for subsequent analysis.

[0047] S21 Clustering, using the K-means clustering algorithm to cluster the historical operation data under the i-th operation index. K-means clustering is a commonly used unsupervised learning algorithm, and its goal is to divide the data set into K different clusters, so that the data points within the same cluster have a high similarity, and the data points between different clusters have a low similarity. In this step, this algorithm is used for the historical operation data under the i-th operation index, and K = 2 is set, that is, the historical operation data under the i-th operation index is divided into two clusters.

[0048] S22 Determine the threshold value. By using the large law method, maximize the between-class variance of the two clusters. The between-class variance reflects the degree of separation between the two clusters. The larger the between-class variance, the more obvious the difference between the two clusters. When the historical operation data that maximizes the between-class variance is found through the large law method, this historical operation data value is determined as the alarm threshold of the i-th operation index.

[0049] S23 Calculate the average value. For each operation index, such as the flow rate and vibration frequency of the device, add up all its historical operation data and then divide by the number of data to obtain the average value of this operation index.

[0050] S24 Data analysis. Adopt the method of class PDP. For the i-th operation index, combine each historical operation data included in the i-th operation index with the average values of the remaining operation indexes respectively to form multiple input samples. For example, if there are three operation indexes: temperature, pressure, and flow rate, when analyzing the temperature index, combine each temperature historical data with the average values of pressure and flow rate respectively, and then input them into the neural network model to obtain the change amount of the prediction result of the i-th operation index. The calculation model of the change amount of the prediction result of the i-th operation index is as follows: ; where, is the output probability when the j-th historical operation data in the i-th operation index is input; is the output probability of the k-th remaining operation index when the j-th historical operation data in the i-th operation index is input; n is the number of historical operation data in the i-th operation index; m is the number of operation indexes.

[0051] The method of class PDP (Partial Dependence Plot) is a method for explaining machine learning models. It can show the influence of a single feature on the model prediction result. In this step, the method of class PDP is applied to the neural network model S25 Obtain the change trend. Obtain the change amounts of all key operation indexes, and set the priorities of the operation indexes according to the order of the change amounts from large to small. The larger the change amount, the more significant the influence of this operation index on the model prediction result, so a higher priority is assigned to it.

[0052] For example, if the change amounts of the two operation indexes of temperature and pressure are 0.3 and 0.1 respectively, then the priority of temperature is higher than that of pressure, and the temperature index will be given priority in subsequent analysis and judgment.

[0053] S26 Threshold judgment. According to the priority of the operation indicators set according to the change trend obtained in S25, check in turn whether the data of each key operation indicator at the current moment exceeds its corresponding alarm threshold. For example, if the temperature has the highest priority, first judge whether the temperature data at the current moment is greater than the alarm threshold of the temperature indicator. If the data of a certain key operation indicator at the current moment exceeds its alarm threshold, it indicates that there is an abnormal situation in the operation indicator of the device. At this time, a data unqualified signal is sent to remind relevant personnel to pay attention to the operation status of the device. If the data of all key operation indicators at the current moment do not exceed the corresponding alarm thresholds, it indicates that the operation status of the device is normal. At this time, S13 modeling and training can be continued.

[0054] In this embodiment, first, the K-means clustering algorithm (K = 2) is used to cluster the historical operation data under the i-th operation indicator to obtain clusters. Then, the between-class variance of the two clusters is maximized by the large law method, and the historical operation data corresponding to the maximum between-class variance is determined as the alarm threshold of the i-th operation indicator. At the same time, the average value of the historical operation data under each operation indicator is calculated respectively. Then, the class PDP method is used to input the historical operation data under the i-th operation indicator and the average values corresponding to the remaining operation indicators into the neural network model in turn to obtain the change amount of the prediction result of the i-th operation indicator. After that, Pareto analysis is performed on the change amounts of each operation indicator, the key operation indicators are screened out and the operation indicator set is updated, and then the priorities of the operation indicators are set from large to small according to the change amounts. Finally, it is judged in turn whether each real-time operation data is greater than the corresponding alarm threshold according to the priority. If it is greater, a data unqualified signal is sent. If not, the steps of modeling and training are executed.

[0055] Example 3: Refer to Figure 3 , the difference between this embodiment and Embodiment 1 is that after executing S14 to predict the corrosion rate and before executing S15 to calculate the remaining life, it further includes: S31 Environment acquisition. Solvent environment data and corrosion current data with the same time stamp as the historical operation data are acquired. The historical operation data records the operation status of the device at different past moments, while the solvent environment data and corrosion current data with the same time stamp can reflect the solvent environment conditions inside the device and the corresponding corrosion current situation under this operation status.

[0056] Acquire the solvent environment data at the current moment. The solvent environment data at the current moment reflects the actual environmental conditions of the current solvent. The solvent environment data of the carbon dioxide capture device is monitored in real time, such as total amine concentration, carbamate concentration, carbonate concentration, amine solution pH value, heat stable salt concentration (HSS concentration), CO2 loading rate (α value), etc. These data will be used to be input into the trained model later to predict the corrosion current at the current moment.

[0057] S32 constructs a dataset, integrates all solvent environment data into a solvent environment dataset, randomly selects solvent environment data from the integrated solvent environment dataset, and constructs p - 1 sub - datasets. During the extraction process, each solvent environment data has the opportunity to be placed in any of the sub - datasets. The above - mentioned extraction method can improve the diversity and randomness of the sub - datasets.

[0058] To further increase the randomness and representativeness of the sub - datasets, the solvent environment dataset is divided into q layers. Then a random number x from 1 to q is randomly generated. A solvent environment data is randomly selected from the corresponding x - th layer and placed in any of the sub - datasets, and then the selected solvent environment data is put back into the solvent environment dataset. This method of stratified random sampling can enable the sub - datasets to better cover various features in the solvent environment dataset.

[0059] For example, if the solvent environment dataset is divided into 5 layers according to a certain feature (such as total amine concentration) and the randomly generated number x is 3, then a solvent environment data is randomly selected from the 3 - rd layer and placed in the sub - dataset.

[0060] S33 Iteration: After each construction of the sub - dataset, the previously selected solvent environment data is put back into the solvent environment dataset to restore the solvent environment dataset to its original state. Then S32 is executed again to construct the dataset, and random extraction and distribution are performed again to generate new sub - datasets. The above - mentioned iterative process is continuously repeated until the number of samples in each sub - dataset exceeds the number of samples in the solvent environment dataset. Finally, all the generated sub - datasets are sequentially updated as the solvent environment dataset.

[0061] S34 Modeling: A random forest model is established. Random forest is an ensemble learning algorithm composed of multiple decision trees. Each decision tree independently performs classification or regression prediction on the data, and finally improves the accuracy and stability of the model by integrating the results of multiple decision trees.

[0062] The p - th solvent environment dataset constructed is used to train the p - th decision tree in the random forest model. During the training process, the decision tree learns the mapping relationship between the data according to the features and labels (if any) in the solvent environment dataset, so as to be able to predict new solvent environment data. For example, the features in the solvent environment dataset include total amine concentration, carbamate concentration, carbonate concentration, amine solution pH value, heat - stable salt concentration (HSS concentration), CO2 loading rate (α value), etc., and the label is the corresponding corrosion current data. The decision tree constructs the corresponding decision rules by learning the relationship between these features and labels.

[0063] After training all decision trees, this embodiment can obtain a trained random forest model. The trained random forest model has learned the complex relationship between solvent environment data and corrosion current and can be used to predict the corrosion current for new solvent environment data.

[0064] S35 Predict the corrosion current. Input the solvent environment data at the current moment obtained into the trained random forest model. The random forest model will analyze and process the current solvent environment data according to the relationship between the previously learned solvent environment data and the corrosion current. The random forest model, based on the input solvent environment data, through the comprehensive judgment of multiple decision trees, outputs the predicted corrosion current data.

[0065] S36 Calculate the corrosion rate. Calculate the theoretical corrosion rate through Faraday's law. The calculation model is as follows: ; Where, is the theoretical corrosion rate; I is the predicted corrosion current data; E is the chemical equivalent of the metal; N is the valence of the metal type used in the carbon dioxide capture device; F is the Faraday constant, with a value of 96485; R is the surface area of the metal, obtained through the design parameters of the carbon dioxide capture device or actual measurement.

[0066] Calculate the corrosion rate difference based on the theoretical corrosion rate and the predicted corrosion rate.

[0067] S37 Corrosion rate judgment. Judge whether the predicted corrosion rate difference is greater than the preset corrosion rate threshold. If so, send an alarm signal; if not, execute S15 to calculate the remaining life.

[0068] This embodiment first obtains the solvent environment data and corrosion current data with the same time stamp as the historical operation data and the solvent environment data at the current moment, then integrates all the solvent environment data into a data set and constructs p - 1 sub - data sets. Through stratified random sampling and continuous iteration until the sample number of the sub - data set is greater than the original solvent environment data set, then update the sub - data set as the solvent environment data set; then establish a random forest model containing p decision trees, and use the p - th solvent environment data set to train the p - th decision tree to obtain the trained model; then input the current solvent environment data into the model to output the predicted corrosion current data, calculate the theoretical corrosion rate through Faraday's law and obtain the difference from the predicted value; finally, judge whether the corrosion rate difference is greater than the preset threshold. If it is greater, send an alarm signal, and if it is less than or equal, execute the step of calculating the remaining life.

[0069] Embodiment 4: This embodiment discloses a corrosion monitoring system for a carbon dioxide capture device. The system includes: The monitoring module includes a real-time data acquisition unit and a historical data acquisition unit.

[0070] The real-time data acquisition unit is used to collect data in real time through a sensor network (such as temperature sensors, pressure sensors, corrosion sensors, etc.), and obtain the real-time operation data, operation time, and corrosion degree before operation of the current carbon dioxide capture device.

[0071] The historical data acquisition unit is used to obtain the final corrosion degree and training data of the same type of carbon dioxide capture device. The training data includes: historical operation data under the full life cycle of the same type of carbon dioxide capture device, and the corrosion rate of the same type of carbon dioxide capture device under each historical operation data.

[0072] The edge computing module includes a modeling and training unit, a predicted corrosion rate unit, and a remaining life calculation unit; The modeling and training unit is used to establish a corrosion prediction model using machine learning algorithms (such as neural networks, random forests, support vector machines, etc.), and train the neural network model with the training data to obtain a trained neural network model for subsequent corrosion rate prediction.

[0073] The predicted corrosion rate unit is used to input the real-time operation data into the trained neural network model and output the predicted corrosion rate.

[0074] The remaining life calculation unit is used to calculate the remaining life based on the final corrosion degree, the corrosion degree before operation, the operation time, and the predicted corrosion rate.

[0075] The system further includes: a memory, and a computer-readable storage medium and training data are stored in the memory.

[0076] A processor, and when the processor calls the program code in the memory, it executes the steps of the method.

[0077] The above are all preferred embodiments of the present application, and the protection scope of the present application is not limited by this. Therefore, all equivalent changes made according to the structure, shape, and principle of the present application should be covered within the protection scope of the present application.

Claims

1. A method for corrosion monitoring of a carbon dioxide capture device, characterized in that, Including: Real-time data acquisition: Obtain the real-time operation data, operation time, and corrosion degree before operation of the current carbon dioxide capture device; Historical data acquisition: Obtain the final corrosion degree and training data of the same type of carbon dioxide capture device. The training data includes: historical operation data under the full life cycle of the same type of carbon dioxide capture device, and the corrosion rate of the same type of carbon dioxide capture device under each historical operation data; Modeling and training: Use a machine learning algorithm to establish a neural network model, and train the neural network model with the training data to obtain a trained neural network model; Predict the corrosion rate: Input the real-time operation data into the trained neural network model, and output the predicted corrosion rate; Calculate the remaining life: Calculate the remaining life based on the final corrosion degree, the corrosion degree before operation, the operation time, and the predicted corrosion rate.

2. The method for corrosion monitoring of the carbon dioxide capture device according to claim 1, wherein The historical operation data includes m types of operation indicators. After performing the step of historical data acquisition and before performing the step of modeling and training, it further includes: Clustering: Use the K-means clustering algorithm to cluster the historical operation data under the i-th type of operation indicator to obtain clusters, where the value of K is 2; Determine the threshold: Maximize the between-class variance of the two clusters by the large law method, and record the historical operation data corresponding to the maximum between-class variance as the alarm threshold of the i-th type of operation indicator; Threshold judgment: Sequentially judge whether each real-time operation data is greater than the corresponding alarm threshold. If so, send a data unqualified signal; if not, perform the step of modeling and training.

3. The method for corrosion monitoring of the carbon dioxide capture equipment according to claim 2, characterized in that, After performing the step of determining the threshold and before performing the step of threshold judgment, it further includes: Calculate the average value: Calculate the average value of the historical operation data under each type of operation indicator respectively; Data analysis: Using a method similar to the PDP method, the historical operation data under the i-th operation index and the average values corresponding to the remaining operation indexes are sequentially input into the neural network model to obtain the change amount of the prediction result of the i-th operation index. The calculation model of the change amount of the prediction result of the i-th operation index is as follows: ； Among them, is the output probability when the j-th historical operation data in the i-th operation index is input; is the output probability of the k-th remaining operation index when the j-th historical operation data in the i-th operation index is input; n is the number of historical operation data in the i-th operation index; m is the number of operation indexes; Obtain the change trend: Obtain all the change amounts, and set the priority of the operation indicators according to the order of the change amounts from large to small; In the step of threshold judgment, compare the real-time operation data with the corresponding alarm threshold according to the priority.

4. The carbon dioxide capture equipment corrosion monitoring method according to claim 3, wherein The method further includes: Indicator screening: Perform Pareto analysis on the change amount of each operation indicator to obtain the key operation indicators, and update the key operation indicators as the operation indicators.

5. The method for corrosion monitoring of a carbon dioxide capture device according to any one of claims 1-4, characterized in that, After performing the step of predicting the corrosion rate and before performing the step of calculating the remaining life, it further includes: Environment acquisition: Obtain the solvent environment data and corrosion current data with the same timestamp as the historical operation data; obtain the solvent environment data at the current moment; Modeling: Establish a random forest model, and train the random forest model with the solvent environment data and corrosion current data to obtain a trained random forest model; Predict the corrosion current: Input the solvent environment data at the current moment into the trained random forest model, and output the predicted corrosion current data; Calculate the corrosion rate: Calculate the theoretical corrosion rate through Faraday's law, and calculate the corrosion rate difference based on the theoretical corrosion rate and the predicted corrosion rate; Corrosion rate judgment: Judge whether the predicted corrosion rate difference is greater than the preset corrosion rate threshold. If so, send a warning signal; if not, perform the step of calculating the remaining life.

6. The corrosion monitoring method of the carbon dioxide capture device according to claim 5, characterized in that, The random forest model includes p decision trees. After the step of obtaining the execution environment and before the step of modeling, it further includes: Constructing a data set: Integrating all solvent environment data into a solvent environment data set, constructing p - 1 sub - data sets, and randomly taking one solvent environment data and putting it into any one of the sub - data sets; Iteration: Putting the selected solvent environment data back into the solvent environment data set and re - executing the step of constructing the data set until the number of samples in all sub - data sets is greater than the number of samples in the solvent environment data set, and updating all sub - data sets as the solvent environment data set; In the step of modeling, the p - th solvent environment data set is used to train the p - th decision tree in the random forest model.

7. The method for corrosion monitoring of carbon dioxide capture equipment according to claim 6, characterized in that, The step of constructing the data set further includes: Dividing the solvent environment data set into q layers, randomly generating a random number x from 1 to q, and randomly taking one solvent environment data from the x - th layer and putting it into any one of the sub - data sets.

8. The carbon dioxide capture equipment corrosion monitoring method according to any one of claims 1-4, characterized in that, The method further includes: Remaining life judgment: Judging whether the remaining life is less than a preset life threshold. If so, execute the warning step; if not, re - execute the step of real - time data collection; Warning: Sending out a warning signal.

9. A corrosion monitoring system for a carbon dioxide capture device, characterized in that, The system is used to execute the method according to any one of claims 1 - 8. The system includes: A monitoring module, including a real - time data collection unit and a historical data collection unit; The real - time data collection unit is used to obtain the real - time operation data, operation time, and corrosion degree before operation of the current carbon dioxide capture device; The historical data collection unit is used to obtain the final corrosion degree and training data of the same - type carbon dioxide capture device. The training data includes: historical operation data of the same - type carbon dioxide capture device throughout its life cycle, and the corrosion rate of the same - type carbon dioxide capture device under each historical operation data; An edge computing module, including a modeling and training unit, a predicted corrosion rate unit, and a remaining life calculation unit; The modeling and training unit is used to establish a neural network model using a machine learning algorithm, and train the neural network model with the training data to obtain a trained neural network model; The predicted corrosion rate unit is used to input the real - time operation data into the trained neural network model and output the predicted corrosion rate; The remaining life calculation unit is used to calculate the remaining life based on the final corrosion degree, the corrosion degree before operation, the operation time, and the predicted corrosion rate.

10. The carbon dioxide capture equipment corrosion monitoring system according to claim 9, wherein, The system further includes: A memory, in which a computer - readable storage medium and training data are stored; A processor, when the processor calls the program code in the memory, executes the steps of the method according to any one of claims 1 - 8.

Images