A method and apparatus for monitoring liquid sulfur blockage in a sulfur recovery unit

Abstract

The present application relates to the technical field of industrial process monitoring, and particularly relates to a sulfur recovery device liquid sulfur blockage monitoring method and equipment, comprising: obtaining real-time data containing process parameters and key pressure parameters; using a time series large model to extract time sequence characteristics, inputting the time sequence characteristics into a time sequence cluster after being fused with the process parameters to obtain probability labels of each working condition category; inputting the process parameters into multiple regressors in parallel to obtain multiple predicted values of the key pressure parameters, performing weighted fusion based on the probability labels to obtain weighted regression predicted values; and determining liquid sulfur blockage according to the deviation of the weighted regression predicted values from true values. The present application can automatically adapt to multiple working condition changes, identify working conditions through time sequence clustering and output probability distribution, effectively distinguish working condition fluctuations from real faults, fuse multiple regressor outputs based on probability weighted regression prediction, improve the prediction robustness for unknown working conditions, and finally realize early and accurate identification of liquid sulfur blockage by analyzing the prediction deviation, thereby significantly improving the equipment adaptability and early warning capability.

Description

Technical Field

[0001] This invention relates to the field of industrial process monitoring technology, and in particular to a method and equipment for monitoring liquid sulfur blockage in a sulfur recovery device. Background Technology

[0002] In the field of liquid sulfur blockage monitoring in sulfur recovery units, existing technologies face two major challenges. Firstly, distributed control systems (DCS) rely on fixed thresholds for alarms, lacking sufficient sensitivity to slowly changing faults like liquid sulfur blockage, making early warning difficult. Secondly, traditional machine learning methods, whether unsupervised learning (such as PCA and SVDD) lacking the ability to capture temporal dynamics and thus insensitive to weak fault signals, or supervised learning suffering from poor model generalization and overfitting due to scarce fault samples and labeling difficulties, are all ill-suited to the complex and ever-changing operating conditions in industrial processes. These methods generally lack dynamic adaptability to multiple operating conditions, easily misinterpreting normal fluctuations as faults or failing to accurately identify true anomalies when conditions change, leading to frequent model updates and low monitoring reliability.

[0003] Therefore, there is an urgent need to provide a technical solution to address the above problems. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method and device for monitoring liquid sulfur blockage in a sulfur recovery device.

[0005] In a first aspect, the present invention provides a method for monitoring liquid sulfur blockage in a sulfur recovery device, the technical solution of which is as follows:

[0006] Acquire real-time data for each operating condition category in the sulfur recovery unit at the current moment; the real-time data includes process parameters and key pressure parameters; use a pre-constructed time series large model to obtain the time-series characteristics of the process parameters;

[0007] The process parameters and the time-series features are fused, and the fused features are input into a pre-trained time-series clusterer to obtain probability labels that characterize the likelihood of the fused features belonging to each working condition category.

[0008] The process parameters are input in parallel to multiple pre-trained regressors to obtain multiple predicted values ​​of the key pressure parameters. Based on the probability labels, the multiple predicted values ​​are weighted and fused to obtain the weighted regression predicted value of the key pressure parameters at the current time. Each regressor corresponds to a working condition category, and each regressor outputs a set of predicted values ​​of key pressure parameters.

[0009] Based on the deviation between the weighted regression predicted value and the actual value of the key pressure parameter, if the deviation of any key pressure parameter exceeds its corresponding preset deviation threshold, it is determined that liquid sulfur blockage has occurred.

[0010] The beneficial effects of the liquid sulfur blockage monitoring method for a sulfur recovery device of the present invention are as follows:

[0011] The method of this invention can automatically adapt to changes in multiple operating conditions. It identifies operating conditions and outputs probability distributions through time-series clustering, effectively distinguishing between operating condition fluctuations and actual faults. Based on probability-weighted regression prediction, it integrates the outputs of multiple regressors, improving the robustness of predictions for unknown operating conditions. Finally, by analyzing prediction deviations, it achieves early and accurate identification of liquid sulfur blockage, significantly improving equipment adaptability and early warning capabilities.

[0012] Based on the above scheme, the liquid sulfur blockage monitoring method of the sulfur recovery device of the present invention can be further improved as follows.

[0013] In one alternative approach, the process parameters include at least one of the following parameters:

[0014] Acid gas flow rate, mixed gas flow rate, combustion air flow rate, combustion air temperature, combustion air pressure, natural gas flow rate, natural gas flow rate and temperature, natural gas flow rate and pressure, fuel gas flow rate, fuel gas temperature, and fuel gas pressure;

[0015] The key pressure parameter includes at least one of the following parameters:

[0016] Main furnace pressure, first-stage Claus outlet pressure, and second-stage Claus outlet pressure.

[0017] In the above-mentioned optional methods, by explicitly listing the specific types of the process parameters and key pressure parameters, the technical solution is closely linked to the specific industrial scenario of the sulfur recovery device, ensuring that the constructed model can accurately monitor the physical parameters most relevant to liquid sulfur blockage, thereby improving the practical application value and pertinence of the solution.

[0018] In one alternative approach, an offline modeling step is also included:

[0019] The missing values ​​in the historical normal operating data of the sulfur recovery unit are filled in layers to obtain the filled historical data. The outliers in the filled historical data are then intelligently corrected to obtain the preprocessed historical data.

[0020] An automated machine learning framework is used to select the optimal model from multiple time series models and train it to construct the large time series model for extracting the time series features.

[0021] Using the trained time series large model, time series features are extracted from the preprocessed historical data, and the extracted time series features are fused with the preprocessed historical data to form historical fusion features;

[0022] Unsupervised clustering is performed on the historical fusion features to divide them into multiple working condition categories in order to complete the training of the temporal clusterer;

[0023] Based on the operating condition categories divided by the time-series clusterer of the preprocessed historical data, a regressor is trained for each operating condition category; wherein, the regressor adopts a multilayer perceptron structure and is used to predict key pressure parameters based on the input process parameters.

[0024] Based on the prediction errors of multiple regressors on the historical normal operating data, the preset deviation thresholds corresponding to different key pressure parameters are determined by the 3σ rule.

[0025] Among the aforementioned optional approaches, by constructing a complete and detailed offline modeling process, covering the entire process from data cleaning, feature extraction, and operating condition classification to model training and threshold setting, it is ensured that online monitoring is based on a benchmark that can comprehensively and accurately reflect the historical normal operating status of the device. This process automatically identifies operating conditions through unsupervised clustering, reducing the reliance on expert knowledge and fault data. Furthermore, the use of the 3σ rule to set thresholds makes the anomaly criteria statistically significant, laying a solid foundation for reliable and accurate anomaly identification in the online phase and significantly improving the systematicness and automation level of the entire method.

[0026] In one alternative approach, the steps of acquiring real-time data for each operating condition category in the sulfur recovery unit at the current moment and obtaining the time-series characteristics of the process parameters using a pre-built time-series large model include:

[0027] The missing values ​​in the initial real-time data of the sulfur recovery device at the current moment are filled in layers to obtain the filled initial real-time data. The outliers in the filled initial real-time data are then intelligently corrected to obtain the real-time data.

[0028] Using the pre-built time series model, time-series features representing long-term and short-term dependencies, local mutations, and multivariate associations are extracted from the real-time data.

[0029] In the above-mentioned optional methods, the online data undergoes the same adaptive preprocessing and temporal feature extraction steps as the offline stage, ensuring that the data caliber is consistent between the online processing and offline modeling stages, and avoiding model performance degradation due to data quality issues. At the same time, by extracting temporal features that can characterize process dynamics (such as long-term and short-term dependencies, local mutations), richer and deeper information is provided for subsequent working condition identification, thereby enabling a better understanding and response to complex temporal changes in the actual production process, and enhancing the real-time performance and accuracy of monitoring.

[0030] In one alternative approach, the step of fusing the process parameters and the time-series features, and inputting the fused features into a pre-trained time-series clusterer to obtain probability labels characterizing the likelihood that the fused features belong to each operating condition category, includes:

[0031] The timing features are combined with the process parameters to form the fused features;

[0032] The fused features are input into the pre-trained temporal clusterer, which outputs a probability distribution. The probability distribution represents the likelihood that the fused features belong to each working condition category, and serves as the probability label.

[0033] In the above-mentioned optional methods, by concatenating and fusing the extracted temporal features with the original process parameters, and using a pre-trained temporal clusterer to output the probability distribution, the feature dimensions of the operating condition discrimination are enriched. This makes the operating condition recognition not only dependent on instantaneous process values, but also considers its dynamic change patterns, thereby capturing the current operating state more delicately and accurately. Outputting probability labels instead of hard categories provides a continuous and flexible weight basis for subsequent weighted fusion, enabling the model to have better interpretability and adaptability for operating states that are in the transition zone of operating conditions or slightly deviate from the training operating conditions.

[0034] In one alternative approach, the step of weightedly fusing multiple predicted values ​​based on the probability labels to obtain the weighted regression predicted value of the key pressure parameter at the current moment includes:

[0035] The probability labels are normalized to obtain the normalized weights corresponding to each working condition category.

[0036] Based on the predicted values ​​output by each regressor and the normalized weights corresponding to each operating condition category, the weighted regression predicted values ​​of the corresponding key pressure parameters at the current moment are determined.

[0037] Among the above-mentioned optional methods, the normalized probability labels are used as weights to perform weighted fusion of the prediction results of all regressors. This strategy creatively solves the problem that a single regressor is difficult to adapt to multiple operating conditions and may encounter unseen operating conditions online. This makes the final prediction value no longer the output of a specific operating condition model, but the consensus or weighted average of the predictions of all known operating condition models. This improves the model's predictive robustness and generalization ability when facing operating condition fluctuations or unforeseen operating states, and effectively prevents missed or false alarms caused by misjudgment of operating conditions.

[0038] In one alternative approach, the step of determining the deviation of the corresponding key pressure parameter based on the weighted regression predicted value and the actual value of the key pressure parameter includes:

[0039] The deviation is determined based on the absolute error between the weighted regression prediction value of the corresponding key pressure parameter at the current moment and the actual value of the corresponding key pressure parameter.

[0040] Among the above-mentioned optional methods, the absolute error is used to calculate the deviation between the weighted regression predicted value and the true value. This method is intuitive and easy to calculate, which is conducive to online monitoring equipment to perform fast and efficient real-time calculation and judgment. The simple deviation calculation logic reduces the complexity of equipment implementation, while providing a clear and direct criterion for comparison with the threshold, ensuring that it can maintain high detection sensitivity and fast response capability for fault modes such as liquid sulfur blockage, which are characterized by continuous deviation of pressure parameters from normal.

[0041] Secondly, the present invention provides a liquid sulfur blockage monitoring device for a sulfur recovery unit, the technical solution of which is as follows:

[0042] The liquid sulfur blockage monitoring equipment for the sulfur recovery unit includes:

[0043] The feature extraction unit is used to acquire real-time data of each operating condition category in the sulfur recovery unit at the current moment; the real-time data includes process parameters and key pressure parameters; and the time-series features of the process parameters are obtained using a pre-built time series large model.

[0044] The working condition identification unit is used to fuse the process parameters and the time-series features, and input the fused features into a pre-trained time-series clusterer to obtain a probability label characterizing the possibility that the fused features belong to each working condition category.

[0045] The weighted prediction unit inputs the process parameters in parallel to multiple pre-trained regressors to obtain multiple predicted values ​​of the key pressure parameters. Based on the probability labels, it performs weighted fusion of the multiple predicted values ​​to obtain the weighted regression prediction value of the key pressure parameters at the current time. Each regressor corresponds to a working condition category, and each regressor outputs a set of predicted values ​​of key pressure parameters.

[0046] The anomaly detection unit determines that liquid sulfur blockage has occurred if the deviation of any key pressure parameter exceeds its corresponding preset deviation threshold, based on the deviation between the weighted regression predicted value and the actual value of the key pressure parameter.

[0047] The beneficial effects of the liquid sulfur blockage monitoring device of the present invention are as follows:

[0048] The device of this invention can automatically adapt to changes in multiple operating conditions. It identifies operating conditions and outputs probability distributions through time-series clustering, effectively distinguishing between operating condition fluctuations and actual faults. Based on probability-weighted regression prediction, it integrates the outputs of multiple regressors, improving the robustness of predictions for unknown operating conditions. Finally, by analyzing prediction deviations, it achieves early and accurate identification of liquid sulfur blockage, significantly improving the adaptability and early warning capabilities of the device.

[0049] Based on the above solution, the liquid sulfur blockage monitoring device of the sulfur recovery device of the present invention can be further improved as follows.

[0050] In one alternative approach, the process parameters include at least one of the following parameters:

[0051] Acid gas flow rate, mixed gas flow rate, combustion air flow rate, combustion air temperature, combustion air pressure, natural gas flow rate, natural gas flow rate and temperature, natural gas flow rate and pressure, fuel gas flow rate, fuel gas temperature, and fuel gas pressure;

[0052] The key pressure parameter includes at least one of the following parameters:

[0053] Main furnace pressure, first-stage Claus outlet pressure, and second-stage Claus outlet pressure.

[0054] In an alternative approach, an offline modeling unit is also included, the offline modeling unit being used for:

[0055] The missing values ​​in the historical normal operating data of the sulfur recovery unit are filled in layers to obtain the filled historical data. The outliers in the filled historical data are then intelligently corrected to obtain the preprocessed historical data.

[0056] An automated machine learning framework is used to select the optimal model from multiple time series models and train it to construct the large time series model for extracting the time series features.

[0057] Using the trained time series large model, time series features are extracted from the preprocessed historical data, and the extracted time series features are fused with the preprocessed historical data to form historical fusion features;

[0058] Unsupervised clustering is performed on the historical fusion features to divide them into multiple working condition categories in order to complete the training of the temporal clusterer;

[0059] Based on the operating condition categories divided by the time-series clusterer of the preprocessed historical data, a regressor is trained for each operating condition category; wherein, the regressor adopts a multilayer perceptron structure and is used to predict key pressure parameters based on the input process parameters.

[0060] Based on the prediction errors of multiple regressors on the historical normal operating data, the preset deviation thresholds corresponding to different key pressure parameters are determined by the 3σ rule.

[0061] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0062] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0063] Figure 1 This is a flowchart illustrating a method for monitoring liquid sulfur blockage in a sulfur recovery device according to the present invention.

[0064] Figure 2 This is a schematic diagram of the overall process of a liquid sulfur blockage monitoring method for a sulfur recovery device according to the present invention;

[0065] Figure 3 This is a schematic diagram of the structure of a liquid sulfur blockage monitoring device for a sulfur recovery apparatus according to the present invention. Detailed Implementation

[0066] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0067] Figure 1 The diagram illustrates a flowchart of an embodiment of a liquid sulfur blockage monitoring method for a sulfur recovery device provided by the present invention. Each step of the method is executed by the liquid sulfur blockage monitoring equipment of the sulfur recovery device, such as... Figure 1 As shown, it includes the following steps:

[0068] S1. Obtain real-time data for each operating condition category in the sulfur recovery unit at the current moment; real-time data includes process parameters and key pressure parameters; use a pre-built time series model to obtain the time-series characteristics of the process parameters. In S1:

[0069] 1) Process parameters refer to input parameters that reflect the operating status of the sulfur recovery unit, including acid gas flow rate, mixed gas flow rate, combustion air flow rate, temperature and pressure, etc.

[0070] 2) Key pressure parameters refer to the output pressure parameters used to directly monitor the liquid sulfur blockage, including the main furnace pressure, the first-stage Claus outlet pressure, and the second-stage Claus outlet pressure.

[0071] 3) The large time series model refers to the optimal model selected from various time series models such as LSTM, GRU, TCN, Transformer, and Informer through an automated machine learning framework. These various time series models cover different paradigms such as recurrent, convolutional, and attention models, and are used to extract deep time series patterns from process parameters.

[0072] 4) Temporal features refer to the features extracted from preprocessed real-time data that can characterize dynamic patterns such as long-term and short-term dependencies, local mutations, and multivariate associations.

[0073] Specifically, real-time operating data of the sulfur recovery unit is first collected from the distributed control system (DCS), which includes process parameters and key pressure parameters. Then, a pre-trained time series model is used to analyze the process parameters and extract time series features that can characterize the dynamic changes of the process, such as long-term and short-term dependence, local mutations and multivariate correlations, to provide information support for subsequent operating condition identification.

[0074] S2. The process parameters and time-series features are fused, and the fused features are input into a pre-trained time-series clusterer to obtain probability labels representing the likelihood that the fused features belong to each operating condition category. In S2:

[0075] 1) Fusion features refer to the comprehensive feature vector formed by splicing process parameters with extracted time-series features. It includes both static process attributes and dynamic time-series patterns.

[0076] 2) Temporal clusterers are a type of unsupervised learning model that receives fused features as input and classifies operating conditions accordingly.

[0077] 3) The probability label is a probability distribution vector, and the value of each element represents the probability that the current operating state belongs to a certain working condition category.

[0078] Specifically, a pre-trained time series large model is used to extract time series features reflecting the dynamic characteristics of equipment operation from process parameters. The time series features are then concatenated with process parameters to form fused features, which are then input into an offline-trained time series clusterer. This clusterer outputs a probability distribution, i.e., a probability label, which represents the degree of matching between the current operating state and each operating condition category.

[0079] S3. The process parameters are input in parallel to multiple pre-trained regressors to obtain multiple predicted values ​​of key pressure parameters. Based on probability labels, the multiple predicted values ​​are weighted and fused to obtain the weighted regression predicted value of the key pressure parameters at the current time. Each regressor corresponds to a working condition category, and each regressor outputs a set of predicted values ​​for key pressure parameters. In S3:

[0080] 1) A regressor is a model built using a multilayer perceptron structure to predict all key pressure parameters based on input process parameters. Each regressor outputs a prediction vector containing the predicted values ​​of each key pressure parameter arranged in a fixed order (e.g., [predicted main furnace pressure, predicted first-stage Claus outlet pressure, predicted second-stage Claus outlet pressure]).

[0081] 2) Weighted fusion refers to the process of summing multiple predicted values ​​for each key stress parameter based on probability labels. Specifically, for each key stress parameter, its predicted values ​​across all regressors are treated as a set, and this set is weighted using the same probability labels to obtain an independent weighted regression predicted value for that parameter.

[0082] Specifically, the process parameters are simultaneously input into all pre-trained regressors. Assume the system has N operating condition categories, corresponding to N trained regressors. Then the nth (n=1,2,...,N) regressor will output a prediction vector. ,in , , These represent the predicted values ​​of the regressor for the main furnace pressure, the first-stage Claus outlet pressure, and the second-stage Claus outlet pressure, respectively.

[0083] Simultaneously, probability labels P=[p1,p2,...,pN] are obtained from S2, representing the probability that the current data belongs to each working condition category. These probability labels are then normalized to obtain the normalized weights for each working condition. .

[0084] Subsequently, weighted fusion was performed independently for each key pressure parameter:

[0085] Weighted regression prediction of main furnace pressure = ;

[0086] The weighted regression forecast for first-order Claus export pressure is:

[0087] ;

[0088] The weighted regression forecast for second-order Claus export pressure is:

[0089] ;

[0090] Ultimately, a comprehensive weighted regression prediction vector is obtained.

[0091] =[Weighted main furnace pressure, weighted primary pressure, weighted secondary pressure].

[0092] It should be noted that the above example using three key pressure parameters illustrates the weighted fusion process. This invention is also applicable to scenarios where only one or two key pressure parameters are monitored. When only one key pressure parameter (such as the main furnace pressure) is monitored, the prediction vector output by each regressor... It will degenerate into a scalar. The weighted fusion process is simplified to Its basic principle is exactly the same as that in the multi-parameter case. This design ensures the flexibility and versatility of the method under different monitoring requirements.

[0093] To further illustrate the weighting process, the following example is provided:

[0094] Assume the system has 3 operating condition categories (N=3) and 3 key pressure parameters. For the process parameters at the current moment, the normalized probability weights given by the time-series clusterer are:

[0095] .

[0096] The predicted vectors output by the three regressors are as follows:

[0097] Regressor 1 (corresponding to operating condition category 1): (Unit: kPa);

[0098] Regressor 2 (corresponding to operating condition category 2): ;

[0099] Regressor 3 (corresponding to operating condition category 3): .

[0100] The weighted regression prediction values ​​for each key pressure parameter are calculated as follows:

[0101] The predicted pressure value of the main furnace is:

[0102] ;

[0103] The first-level Claus export pressure forecast is:

[0104] ;

[0105] The second-level Claus export pressure forecast is:

[0106] .

[0107] Therefore, the final weighted regression prediction vector at the current moment is:

[0108] This vector will be used to compare with the actual measurement in S4.

[0109] S4. Based on the deviation between the weighted regression predicted value and the actual value of the key pressure parameter, if the deviation of any key pressure parameter exceeds its corresponding preset deviation threshold, liquid sulfur blockage is determined to have occurred. In S4:

[0110] 1) The true value refers to the actual measured value of the key pressure parameter that is the same as the target predicted by the regressor and is directly obtained from the real-time data collected by the DCS system.

[0111] 2) The preset deviation threshold is the allowable range determined by statistical analysis of the prediction errors of multiple regressors on historical normal operating data. The preset deviation threshold is calculated based on historical normal operating data using the 3σ rule.

[0112] Specifically, the actual measured values ​​of key pressure parameters are obtained from real-time data, i.e., the true values. The absolute error between the true values ​​and the weighted regression predicted values ​​is calculated as the deviation. The deviation is compared with the preset deviation threshold determined in the offline modeling stage. If the deviation continues to exceed the threshold, it is determined that liquid sulfur blockage has occurred, and an alarm signal is issued to remind the operator to deal with it in time.

[0113] The technical solution of this embodiment can automatically adapt to changes in multiple operating conditions. It identifies operating conditions and outputs probability distributions through time-series clustering, effectively distinguishing between operating condition fluctuations and actual faults. The probability-weighted regression prediction integrates the outputs of multiple regressors, improving the robustness of predictions for unknown operating conditions. Finally, by analyzing prediction deviations, it achieves early and accurate identification of liquid sulfur blockage, significantly improving equipment adaptability and early warning capabilities.

[0114] In one alternative approach, the process parameters include at least one of the following parameters:

[0115] Acid gas flow rate, mixed gas flow rate, combustion air flow rate, combustion air temperature, combustion air pressure, natural gas flow rate, natural gas flow rate and temperature, natural gas flow rate and pressure, fuel gas flow rate, fuel gas temperature, and fuel gas pressure;

[0116] Key pressure parameters include at least one of the following:

[0117] Main furnace pressure, first-stage Claus outlet pressure, and second-stage Claus outlet pressure.

[0118] It should be noted that the selection of process parameters and key pressure parameters in this invention is not arbitrary, but determined based on the process mechanism of the Claus sulfur recovery unit and the failure characteristics of liquid sulfur blockage. The specific reasons are as follows:

[0119] The selected process parameters collectively constitute the core input conditions characterizing the operating status of the sulfur recovery unit. The acid gas flow rate directly determines the input load of the reactants and is the foundation of the entire reaction process; the mixed gas flow rate affects the material ratio within the reaction system; the flow rates of combustion air, natural gas, and fuel gas collectively provide the necessary oxygen and energy for the combustion reaction in the main furnace, which is crucial for maintaining stable reaction temperature; the temperature and pressure parameters of the aforementioned media (acid gas, mixed gas, combustion air, natural gas, and fuel gas) accurately reflect their physical state upon entering the unit and are important boundary conditions affecting reaction rate, flow field distribution, and pressure balance. These process parameters, as inputs to the model, comprehensively define the operating setpoints of the unit under different operating conditions.

[0120] The selected key pressure parameters are output variables that directly respond to liquid sulfur blockage faults. The main furnace pressure, the first-stage Claus outlet pressure, and the second-stage Claus outlet pressure are the core pressure monitoring points sequentially passed along the process pipelines in the sour gas treatment process. When liquid sulfur gradually accumulates in pipelines or equipment, forming a blockage, it directly leads to a reduction in the flow cross-section and an increase in resistance, thus manifesting as a sustained abnormal rise in pressure at the aforementioned key nodes. More importantly, these pressure parameter changes exhibit a clear causal transmission relationship; that is, pressure anomalies in upstream equipment directly lead to pressure changes in downstream equipment. Therefore, using these strongly correlated pressure parameters as the model's prediction targets and monitoring objects can most directly and sensitively capture early signals of liquid sulfur blockage.

[0121] Therefore, by selecting the aforementioned specific process parameters as inputs and monitoring the aforementioned specific key pressure parameters as outputs, a dynamic mapping model between "input-output" under normal operating conditions can be accurately established. Once a significant and persistent deviation occurs between the pressure value predicted by the model and the actual measured value, it can be efficiently and accurately determined that liquid sulfur blockage has occurred.

[0122] In one alternative approach, an offline modeling step is also included:

[0123] The missing values ​​in the historical normal operating data of the sulfur recovery unit are filled in layers to obtain the filled historical data. The outliers in the filled historical data are then intelligently corrected to obtain the preprocessed historical data.

[0124] By using an automated machine learning framework, the optimal model is selected from multiple time series models and trained to build a large time series model for extracting time series features;

[0125] Using a well-trained time series model, time series features are extracted from preprocessed historical data, and the extracted time series features are fused with the preprocessed historical data to form historical fused features;

[0126] Unsupervised clustering of historical fusion features is performed to divide them into multiple working condition categories to complete the training of the temporal clusterer;

[0127] Based on the operating condition categories divided by the time series clusterer from the preprocessed historical data, a regressor is trained for each operating condition category; the regressor adopts a multilayer perceptron structure to predict key pressure parameters based on the input process parameters.

[0128] Based on the prediction errors of multiple regressors on historical normal operating data, the preset deviation thresholds corresponding to different key pressure parameters are determined by the 3σ rule.

[0129] In this embodiment, adaptive data preprocessing is performed on historical normal operating condition data to obtain preprocessed historical data. Further:

[0130] Historical normal operating condition data collected by the DCS system is recorded as follows: , where R is the set of real numbers, M represents the number of bit numbers of the corresponding data, and T represents the sequence length of the corresponding data;

[0131] The missing values ​​in the historical data X are filled in a layered manner, including using forward filling for a small number of missing single digits and using an autoencoder to fill for low-quality data with large areas of missing values.

[0132] Intelligent correction of outliers in the filled data includes filtering out abnormal outliers using statistical methods and correcting them based on data from previous and subsequent time points;

[0133] The preprocessed historical data obtained after adaptive data preprocessing is denoted as follows: ;

[0134] In this embodiment, various temporal models include Long Short-Term Memory Network (LSTM), Gated Recurrent Unit (GRU), Temporal Convolutional Network (TCN), Temporal Transformer, and Informer, covering different paradigms such as recurrent, convolutional, and attention-based models.

[0135] The preprocessed historical data were processed using the selected optimal time series model. Perform time-series feature extraction, and denote the extracted time-series features as follows: Where D represents the feature dimension;

[0136] The time series feature Z can characterize the short-term and long-term dependencies, local abrupt changes, long-term trends, and multivariate correlations in industrial process parameters;

[0137] The completed time series model is used to extract discriminative time-series features from real-time data during the online monitoring phase;

[0138] It should be noted that an automated machine learning framework refers to a machine learning system that automatically selects the optimal model from a set of candidate models and optimizes its parameters. By using a model-based automated learning framework (AutoML) to select the optimal temporal model and parameters to obtain temporal features, suboptimal parameter selection caused by manual model selection can be avoided. Long Short-Term Memory (LSTM) networks are recurrent neural networks capable of capturing long-term dependencies in time series data. Temporal Convolutional Networks (TCNs) are one-dimensional convolutional neural networks that utilize causal convolution to process temporal data. Temporal Transformers are neural network architectures based on self-attention mechanisms for processing temporal data.

[0139] The large time series model obtained through the above training process can adaptively extract diverse temporal features from industrial process parameters, providing effective feature representations for subsequent operating condition identification.

[0140] In this embodiment, a pre-trained time series large model is used to process the pre-processed historical data. Extract time-series features, and denote the extracted time-series features as follows: Where D represents the feature dimension, M represents the number of tag numbers of the acquired process parameters, and T represents the sequence length of the acquired process parameters.

[0141] The extracted time-series features Z are compared with the preprocessed historical data. The process of integration resulted in historical integration characteristics:

[0142] ;

[0143] in, This refers to the fused feature formed by concatenating time-series features with preprocessed historical data along the feature dimension.

[0144] Input historical fusion features into the clustering model for unsupervised clustering:

[0145] ;

[0146] in, Indicates the first The process parameters at time 1 are marked as the 1st time. Operating conditions, and , representing a defined set of working condition categories; It is a temporal clusterer based on a large temporal model.

[0147] The temporal clusterer is trained by using multiple operating condition categories determined through an unsupervised clustering process.

[0148] It should be noted that time-series features Z can characterize the long-term and short-term dependencies, local mutations, long-term trends, and multivariate correlations in industrial process parameters; unsupervised clustering refers to a machine learning method that automatically divides groups based on the inherent similarity of data without pre-labeling; and operating condition categories refer to a set of process states with similar operating characteristics.

[0149] The time-series clusterer obtained through the above training process can automatically identify and classify various operating conditions in industrial processes, providing an accurate basis for subsequent multi-condition regression prediction.

[0150] In this embodiment, the preprocessed historical data is analyzed using a time-series clusterer. Classification of working conditions For each working condition category Use the corresponding data subset Train a regressor separately:

[0151] ;

[0152] in, It is the first Regression predictions for similar working conditions It is a regressor.

[0153] The regressor adopts a multilayer perceptron (MLP) structure and completes the regression from input features to predicted values ​​through multilayer nonlinear transformation. It is used to predict key pressure parameters based on the input process parameters. The key pressure parameters include the main furnace pressure, the first-stage Claus outlet pressure, and the second-stage Claus outlet pressure.

[0154] By training all regressors, a correspondence between the working condition categories and the regression prediction model is established.

[0155] It should be noted that the multilayer perceptron (MLP) structure refers to a feedforward neural network that includes an input layer, a hidden layer, and an output layer; the key pressure parameters refer to the core monitoring parameters that reflect the liquid sulfur blockage status of the sulfur recovery unit; the input process parameters include acid gas flow rate and combustion air pressure parameters, but do not include pressure parameters of subsequent stages.

[0156] The multiple regressors obtained through the above training process can establish accurate pressure parameter prediction models for different working conditions, providing a reliable model basis for weighted regression prediction in the online monitoring stage.

[0157] In this embodiment, a preset deviation threshold is determined based on the prediction errors of multiple regressors on historical normal operating condition data:

[0158] The absolute error between the predicted and actual values ​​for each regressor Perform calculations; where, where This represents the true value of the nth sample. This represents the predicted value of the nth sample;

[0159] The standard deviation of the absolute error is calculated using the 3σ rule, and three times the standard deviation is used as the preset deviation threshold, denoted as . .

[0160] It should be noted that the 3σ rule refers to a statistical method based on the assumption of normal distribution, using three standard deviations as the control limit.

[0161] Through the above offline modeling steps, a complete system of working condition identification and regression prediction models was established, and accurate deviation judgment criteria were provided for the online monitoring stage.

[0162] In one alternative approach, the steps of acquiring real-time data for each operating condition category in the sulfur recovery unit at the current moment and using a pre-built time-series large model to obtain the time-series characteristics of the process parameters include:

[0163] The missing values ​​in the initial real-time data of the sulfur recovery device at the current moment are filled in layers to obtain the filled initial real-time data. The outliers in the filled initial real-time data are then intelligently corrected to obtain the real-time data.

[0164] By using pre-built large time series models, time series features representing long-term and short-term dependencies, local mutations, and multivariate associations are extracted from real-time data.

[0165] In this embodiment, the real-time process parameters collected by the DCS system, i.e., the initial real-time data, are recorded as follows: Where M represents the number of bit numbers corresponding to the data. This indicates the sequence length of the corresponding data. Further adaptive data preprocessing is performed based on a pre-built large-scale time series model.

[0166] First, the real-time process parameters The missing values ​​in the data are filled using a layered imputation process, including: forward imputation for a small number of single-digit missing values; and autoencoder imputation for low-quality data with large areas of missing values. The data obtained after the layered imputation process is denoted as [data name missing]. This refers to the initial real-time data after filling.

[0167] Then, the initial real-time data after filling. The system performs intelligent outlier correction, including: quickly filtering out anomalous outliers using statistical methods; and correcting the filtered outliers based on data from previous and subsequent time points. The final preprocessed data obtained after intelligent outlier correction is the real-time data.

[0168] It should be noted that forward imputation refers to the method of filling in missing values ​​with the values ​​of the previous valid time step; autoencoder refers to a neural network model that learns the data distribution characteristics through an encoder-decoder structure and generates reasonable imputation values ​​based on the learned characteristics; previous and next time step data refers to the normal process parameters of the previous and next time steps adjacent to the current outlier data point in the time series.

[0169] Furthermore, in this scheme, the initial real-time data contains two types of data with different properties: process parameters as model input and key pressure parameters as monitoring benchmarks. To ensure the quality of model input, data preprocessing and time-series feature extraction are performed only on process parameters. The true values ​​of key pressure parameters remain in their original state and, together with the preprocessed process parameters, are referred to as real-time data. This data is directly used in subsequent steps to compare deviations with predicted values. This approach aims to avoid any modification to the monitored target variables, thereby accurately reflecting the true pressure state of the unit.

[0170] The adaptive data preprocessing steps described above effectively address various data quality issues in real-time process parameters, such as random missing values, large-area missing values, and outliers, significantly improving data quality and providing a reliable data foundation for subsequent operating condition identification and regression prediction.

[0171] In one alternative approach, the steps of fusing process parameters and time-series features, and inputting the fused features into a pre-trained time-series clusterer to obtain probability labels representing the likelihood that the fused features belong to each operating condition category, include:

[0172] The timing features are combined with process parameters to form a fused feature;

[0173] The fused features are input into a pre-trained temporal clusterer, which outputs a probability distribution. The probability distribution represents the likelihood that the fused features belong to each working condition category, and serves as a probability label.

[0174] In this embodiment, a pre-built large time series model is used to analyze pre-processed real-time data. Temporal features are extracted, and the extracted temporal features are denoted as Z. A pre-built large-scale time series model refers to a model obtained by selecting the optimal model from multiple time series models and training it using an automated machine learning framework during the offline modeling phase.

[0175] The extracted time-series features Z are compared with process parameters. By splicing the pieces together, a fused feature is formed:

[0176] ;

[0177] in, This refers to the fused feature formed by splicing time-series features and process parameters along the feature dimension.

[0178] The fused features are input into a pre-trained temporal clusterer, which outputs a probability distribution:

[0179] ;

[0180] in, It is a pre-trained temporal clusterer. This represents the probability that the current working condition data is classified into the nth working condition category, where N represents the total number of working condition categories. A pre-trained temporal clusterer refers to a model obtained through unsupervised clustering training based on historical fusion features during the offline modeling phase; the probability distribution, i.e., the probability label... The probability of the fused features belonging to each working condition category is represented as a probability label.

[0181] It should be noted that time series features represent time series patterns of long-term and short-term dependencies, local mutations, and multivariate associations; time series clusterers are clustering algorithm models based on large time series models; and probability labels are probability distribution vectors that represent the likelihood of the current data belonging to each working condition category.

[0182] Furthermore, the offline and online phases achieve a transition from operating condition classification to operating condition identification through different applications of the temporal clusterer. During offline training, the equipment concatenates the temporal features of historical data with process parameters to form historical fusion features, which are then input into the temporal clusterer for unsupervised learning. The direct output of the clusterer is the definite operating condition category to which each historical data point belongs. During online monitoring, the temporal clusterer acts as a discriminator for the fusion features formed from real-time data, and its output is a probability distribution vector (e.g., [0.1, 0.8, 0.1]). For consistency, the category classification information obtained from offline clustering and the probability distribution vector obtained from online clustering are collectively referred to as probability labels. The probability labels in the offline phase are implicit and deterministic (e.g., a probability of 1 for one category and 0 for the rest), while the probability labels in the online phase are explicit and continuous probability values. This design allows the model to handle transitional states and unknown operating conditions more precisely, providing important weighting criteria for subsequent weighted regression predictions.

[0183] Through the above process, by utilizing the large time series model and time series clusterer trained in the offline stage, the working condition probability of real-time data is identified, providing an accurate probability weight basis for subsequent weighted regression prediction.

[0184] In one alternative approach, the step of weighted fusion of multiple predicted values ​​based on probability labels to obtain the weighted regression predicted value of the key pressure parameter at the current moment includes:

[0185] The probability labels are normalized to obtain the normalized weights corresponding to each working condition category.

[0186] Based on the predicted values ​​output by each regressor and the normalized weights corresponding to each operating condition category, the weighted regression predicted values ​​of the corresponding key pressure parameters at the current moment are determined.

[0187] In this embodiment, probability labels are used. Normalization is performed, and the overall probability is made to satisfy the following using the Softmax function:

[0188] ;

[0189] in, It is the normalized probability value, representing the normalized weight of the nth type of working condition.

[0190] The process parameters at the current moment The data is fed in parallel into all pre-trained regressors for regression prediction, and the regression result is denoted as:

[0191] ;

[0192] in, Indicates the first The predicted values ​​output by the regressor for each type of working condition. The pre-trained regressor refers to a multilayer perceptron model trained separately for each type of working condition during the offline modeling phase.

[0193] The predictions from each regressor are weighted and fused. The weighted regression prediction results are as follows:

[0194] ;

[0195] in, This represents the weighted regression prediction value of the corresponding key pressure parameter at the current moment.

[0196] Through a weighted fusion process, even if the online data does not belong to a certain fixed working condition category in the training set, reasonable prediction results can be obtained through probability weight allocation.

[0197] It should be noted that the Softmax function is a normalized exponential function that maps a set of real numbers to a probability distribution; the pre-trained regressor includes N regression models corresponding to different working conditions; and the weighted regression prediction is the comprehensive prediction result after considering the probability of each working condition.

[0198] Through the above weighted fusion steps, the advantages of multiple regressors trained in the offline stage are fully utilized. Combined with the uncertainty of online working condition identification, adaptive prediction of unknown working conditions is achieved, which significantly improves the model's generalization ability and practical application value.

[0199] In one alternative approach, the step of determining the deviation of the corresponding key stress parameter based on the weighted regression predicted value and the actual value of the key stress parameter includes:

[0200] The deviation is determined based on the absolute error between the weighted regression prediction value of the corresponding key pressure parameter at the current moment and the actual value of the corresponding key pressure parameter.

[0201] In this embodiment, the actual measured values ​​of the corresponding key pressure parameters are obtained from real-time data and denoted as follows: Key pressure parameters include main furnace pressure, first-stage Claus outlet pressure, and second-stage Claus outlet pressure.

[0202] Calculate the weighted regression predicted value Compared with the true value Absolute error:

[0203] ;

[0204] And the absolute error is taken as the deviation.

[0205] Compare the deviation with the preset deviation threshold The preset deviation threshold is calculated using the 3σ rule based on the prediction errors of multiple regressors on historical normal operating data during the offline modeling stage.

[0206] When the deviation exceeds the preset deviation threshold, that is At that time, it was determined that liquid sulfur blockage had occurred inside the device.

[0207] It should be noted that the actual measured value of the key pressure parameter refers to the pressure sensor reading collected in real time from the DCS system; the absolute error refers to the absolute difference between the predicted value and the actual value; and the preset deviation threshold is the upper limit of the allowable error obtained based on the statistical analysis of historical normal operating data.

[0208] By using the above deviation determination process and combining it with the judgment criteria established in the offline stage, pressure anomalies caused by liquid sulfur blockage can be effectively identified, enabling precise monitoring and early warning of the unit's operating status.

[0209] like Figure 2 The overall processing flow diagram shown is divided into an offline training section and an online testing section, further:

[0210] In the offline training section, historical normal operation data of the sulfur recovery unit is collected from the DCS system, i.e., offline training data. The unprocessed offline training data is used as raw data for data preprocessing. By performing hierarchical imputation of missing values ​​in the raw data and intelligent correction of outliers, high-quality regular data is obtained, i.e., preprocessed data. A pre-built time series large model is used to extract dynamic features that can characterize long-term and short-term dependencies, local mutations, and multivariate associations from the preprocessed data, i.e., time series features. The extracted time series features are concatenated with the preprocessed data (i.e., original operating condition attributes) to form fused features. A multi-operating condition collector, i.e., a time series clusterer, receives the fused features as input and divides the preprocessed data into multiple different operating condition categories through unsupervised clustering. For each operating condition category divided by the time series clusterer, a dedicated prediction model, i.e., a regressor, is trained to predict key pressure parameters. Based on the prediction errors of all regressors on historical normal operating condition data, the 3σ rule is used to calculate and determine the preset deviation threshold.

[0211] In the offline training section, the online test data is the sulfur recovery unit's operating data collected in real time from the DCS system, i.e., the initial real-time data. The unprocessed initial real-time data is also adaptively preprocessed to obtain real-time data. For the real-time data, on the one hand, the offline-trained time series model is used to extract time-series features in real time; on the other hand, the preprocessed real-time data itself is also directly used as online data in subsequent processes. The time-series clusterer receives the fused features formed by the time-series features and the preprocessed real-time data, and outputs a probability label, which represents the probability distribution of the possibility that the fused features belong to each operating condition category. The preprocessed real-time data is input into all regressors in parallel to obtain multiple predicted values, and then a weighted sum is performed according to the weights normalized by the probability label.

[0212] By comparing the deviation between the weighted regression prediction value in the online testing phase and the actual value in the real-time data, and whether it exceeds the preset deviation threshold obtained in the offline phase, it is ultimately determined whether liquid sulfur blockage has occurred.

[0213] To further illustrate the specific implementation of the present invention, an application scenario example based on actual operating data of a sulfur recovery device is provided:

[0214] In this example, a Claus sulfur recovery unit is used as the specific application. This unit is the core equipment for treating acidic gas containing hydrogen sulfide and realizing sulfur resource recovery. The acidic gas in the unit needs to undergo desulfurization through key processes such as combustion in the main furnace and catalytic conversion in two-stage Claus reactors. Among these, the pressure of the main furnace, the outlet pressure of the first-stage Claus reactor, and the outlet pressure of the second-stage Claus reactor are the core monitoring indicators reflecting whether liquid sulfur has become blocked.

[0215] Normal operating data of the sulfur recovery unit from 00:00 on April 1, 2025 to 00:00 on June 30, 2025, i.e., historical normal operating data, was collected at a frequency of 10 seconds, resulting in a historical dataset containing multiple process parameters. The collected input parameters included: acid gas (load) flow rate, mixed gas flow rate, combustion air flow rate, and temperature and pressure parameters of acid gas, mixed gas, combustion air, natural gas, and fuel gas; the output target parameters were the main furnace pressure, the first-stage Claus outlet pressure, and the second-stage Claus outlet pressure. A sample data table is shown below:

[0216] Table 1:

[0217]

[0218] Historical normal operating condition data was divided into training, validation, and test sets in a 6:2:2 ratio. Based on the aforementioned adaptive data preprocessing method, the historical normal operating condition data was cleaned and repaired. A time-series large-scale model was used for operating condition identification and regression prediction modeling. The offline operating conditions were divided into 8 categories using a time-series clusterer. This classification result perfectly matched the events recorded in the field operation log: one shutdown / maintenance restart, one coal ratio replacement, and six combustion air flow adjustments.

[0219] For each operating condition category, a corresponding regressor is trained to establish a mapping relationship from input process parameters to key pressure parameters. Based on the prediction errors of all regressors on historical normal data, the deviation threshold is calculated using the 3σ rule. .

[0220] The measured values ​​of the aforementioned process parameters are collected in real time and processed using the same preprocessing methods as in the offline stage. Temporal features are extracted using a pre-trained large-scale time series model, and a temporal clusterer is used to obtain the probability distribution of the current data belonging to each operating condition category. The real-time data is input in parallel into all pre-trained regressors to obtain multiple predicted values, which are then weighted and fused based on the probability distribution to obtain the final predicted value.

[0221] By comparing the deviation between the weighted regression predicted value and the actual measured value, we can determine whether it exceeds a threshold. This enables early detection of liquid sulfur blockage. When a deviation is detected to be continuously exceeding the limit, the equipment issues an alarm to prompt operators to take timely action.

[0222] Further, detailed steps are as follows:

[0223] ① Online real-time data Missing values ​​were filled and outliers were removed using the same preprocessing method as the offline historical normal operating condition data, denoted as... .

[0224] ② Extract time series features by using the best time series model and optimal parameters recommended by the large time series model. The time-series features are combined with preprocessed real-time data and fed into a trained classifier, i.e., a time-series clusterer. In this example, the time-series clusterer obtains probability labels for the current and historical operating conditions, yielding probability values, denoted as [value missing]. The main furnace pressure prediction based on weighted probability can be expressed by the formula:

[0225] ;

[0226] ③ Compare the difference between the predicted value and the actual value to determine whether... To determine whether liquid sulfur blockage has occurred.

[0227] This example demonstrates the practical application process of the technical solution of the present invention in an industrial field. Through multi-condition identification and weighted prediction strategies, the accuracy and reliability of liquid sulfur blockage monitoring are effectively improved.

[0228] Figure 3 A schematic diagram of an embodiment of a liquid sulfur blockage monitoring device 200 for a sulfur recovery apparatus provided by the present invention is shown. Figure 3 As shown, the device 200 includes:

[0229] The feature extraction unit 210 is used to acquire real-time data of each operating condition category in the sulfur recovery unit at the current moment; the real-time data includes process parameters and key pressure parameters; the time series features of the process parameters are obtained using a pre-built time series large model;

[0230] The working condition identification unit 220 is used to fuse process parameters and time-series features, and input the fused features into a pre-trained time-series clusterer to obtain probability labels representing the likelihood that the fused features belong to each working condition category.

[0231] The weighted prediction unit 230 inputs process parameters in parallel to multiple pre-trained regressors to obtain multiple predicted values ​​of key pressure parameters. Based on probability labels, it performs weighted fusion of multiple predicted values ​​to obtain the weighted regression predicted value of the key pressure parameters at the current time. Each regressor corresponds to a working condition category, and each regressor outputs a set of predicted values ​​of key pressure parameters.

[0232] The anomaly detection unit 240 determines that liquid sulfur blockage has occurred if the deviation of any key pressure parameter exceeds its corresponding preset deviation threshold, based on the deviation between the weighted regression predicted value and the actual value of the key pressure parameter.

[0233] Based on the above solution, the liquid sulfur blockage monitoring device of the sulfur recovery device of the present invention can be further improved as follows.

[0234] In one alternative approach, the process parameters include at least one of the following parameters:

[0235] Acid gas flow rate, mixed gas flow rate, combustion air flow rate, combustion air temperature, combustion air pressure, natural gas flow rate, natural gas flow rate and temperature, natural gas flow rate and pressure, fuel gas flow rate, fuel gas temperature, and fuel gas pressure;

[0236] The key pressure parameter includes at least one of the following parameters:

[0237] Main furnace pressure, first-stage Claus outlet pressure, and second-stage Claus outlet pressure.

[0238] In an alternative approach, an offline modeling unit is also included, the offline modeling unit being used for:

[0239] The missing values ​​in the historical normal operating data of the sulfur recovery unit are filled in layers to obtain the filled historical data. The outliers in the filled historical data are then intelligently corrected to obtain the preprocessed historical data.

[0240] An automated machine learning framework is used to select the optimal model from multiple time series models and train it to construct the large time series model for extracting the time series features.

[0241] Using the trained time series large model, time series features are extracted from the preprocessed historical data, and the extracted time series features are fused with the preprocessed historical data to form historical fusion features;

[0242] Unsupervised clustering is performed on the historical fusion features to divide them into multiple working condition categories in order to complete the training of the temporal clusterer;

[0243] Based on the operating condition categories divided by the time-series clusterer of the preprocessed historical data, a regressor is trained for each operating condition category; wherein, the regressor adopts a multilayer perceptron structure and is used to predict key pressure parameters based on the input process parameters.

[0244] Based on the prediction errors of multiple regressors on the historical normal operating data, the preset deviation thresholds corresponding to different key pressure parameters are determined by the 3σ rule.

[0245] The technical solution of this embodiment can automatically adapt to changes in multiple operating conditions. It identifies operating conditions and outputs probability distributions through time-series clustering, effectively distinguishing between operating condition fluctuations and actual faults. The probability-weighted regression prediction integrates the outputs of multiple regressors, improving the robustness of predictions for unknown operating conditions. Finally, by analyzing prediction deviations, it achieves early and accurate identification of liquid sulfur blockage, significantly improving equipment adaptability and early warning capabilities.

[0246] The parameters and steps for each module in the liquid sulfur blockage monitoring device 200 of the sulfur recovery device described above to achieve their respective functions can be referred to the parameters and steps in the embodiments of the liquid sulfur blockage monitoring method of the sulfur recovery device above, and will not be repeated here.

[0247] Those skilled in the art will recognize that this invention can be implemented as a system, method, or computer program product. Therefore, this disclosure can be specifically implemented in the following forms: it can be entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, generally referred to herein as a "circuit," "module," or "system." Furthermore, in some embodiments, the invention can also be implemented as a computer program product contained in one or more computer-readable media, which includes computer-readable program code.

[0248] Any combination of one or more computer-readable media can be used. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0249] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for monitoring liquid sulfur blockage in a sulfur recovery device, characterized in that, include: Obtain real-time data for each operating condition category in the sulfur recovery unit at the current moment; the real-time data includes process parameters and key pressure parameters; use A pre-constructed time series large model is used to obtain the time series characteristics of the process parameters; wherein, the process parameters are input parameters reflecting the operating status of the sulfur recovery unit, including at least one of acid gas flow rate, mixed gas flow rate, combustion air flow rate, temperature and pressure parameters; the key pressure parameters are output parameters used to directly monitor the liquid sulfur blockage status, including at least one of main furnace pressure, first-stage Claus outlet pressure and second-stage Claus outlet pressure. The process parameters and the time-series features are fused, and the fused features are input into a pre-trained time-series clusterer to obtain probability labels representing the likelihood that the fused features belong to each operating condition category. The fused features are comprehensive feature vectors formed by concatenating the process parameters and the time-series features, and simultaneously include static process attributes and dynamic time-series patterns. The time-series clusterer is a clusterer trained based on an unsupervised learning model, used to classify the fused features according to operating conditions. The process parameters are input in parallel to multiple pre-trained regressors to obtain multiple predicted values ​​of the key pressure parameters. Based on the probability labels, the multiple predicted values ​​are weighted and fused to obtain the weighted regression predicted value of the key pressure parameters at the current time. Each regressor corresponds to a working condition category, and each regressor outputs a set of predicted values ​​of key pressure parameters. Based on the deviation between the weighted regression predicted value and the actual value of the key pressure parameter, if the deviation of any key pressure parameter exceeds its corresponding preset deviation threshold, it is determined that liquid sulfur blockage has occurred.

2. The method for monitoring liquid sulfur blockage in a sulfur recovery device according to claim 1, characterized in that, The process parameters include at least one of the following parameters: Acid gas flow rate, mixed gas flow rate, combustion air flow rate, combustion air temperature, combustion air pressure, natural gas flow rate, natural gas flow rate and temperature, natural gas flow rate and pressure, fuel gas flow rate, fuel gas temperature, and fuel gas pressure; The key pressure parameter includes at least one of the following parameters: Main furnace pressure, first-stage Claus outlet pressure, and second-stage Claus outlet pressure.

3. The method for monitoring liquid sulfur blockage in a sulfur recovery device according to claim 1, characterized in that, It also includes offline modeling steps: The missing values ​​in the historical normal operating data of the sulfur recovery unit are filled in layers to obtain the filled historical data. The outliers in the filled historical data are then intelligently corrected to obtain the preprocessed historical data. An automated machine learning framework is used to select the optimal model from multiple time series models and train it to construct the large time series model for extracting the time series features. Using the trained time series large model, time series features are extracted from the preprocessed historical data, and the extracted time series features are fused with the preprocessed historical data to form historical fusion features; Unsupervised clustering is performed on the historical fusion features to divide them into multiple working condition categories in order to complete the training of the temporal clusterer; Based on the operating condition categories divided by the time-series clusterer of the preprocessed historical data, a regressor is trained for each operating condition category; wherein, the regressor adopts a multilayer perceptron structure and is used to predict key pressure parameters based on the input process parameters. Based on the prediction errors of multiple regressors on the historical normal operating data, the preset deviation thresholds corresponding to different key pressure parameters are determined by the 3σ rule.

4. The method for monitoring liquid sulfur blockage in a sulfur recovery device according to claim 1, characterized in that, The steps of acquiring real-time data for each operating condition category in the sulfur recovery unit at the current moment, and using a pre-built large time series model to obtain the time-series characteristics of the process parameters, include: The missing values ​​in the initial real-time data of the sulfur recovery device at the current moment are filled in layers to obtain the filled initial real-time data. The outliers in the filled initial real-time data are then intelligently corrected to obtain the real-time data. Using the pre-built time series model, time-series features representing long-term and short-term dependencies, local mutations, and multivariate associations are extracted from the real-time data.

5. The method for monitoring liquid sulfur blockage in a sulfur recovery device according to claim 1, characterized in that, The steps of fusing the process parameters and the time-series features, and inputting the fused features into a pre-trained time-series clusterer to obtain probability labels characterizing the likelihood that the fused features belong to each operating condition category, include: The timing features are combined with the process parameters to form the fused features; The fused features are input into the pre-trained temporal clusterer, which outputs a probability distribution. The probability distribution represents the likelihood that the fused features belong to each working condition category, and serves as the probability label.

6. The method for monitoring liquid sulfur blockage in a sulfur recovery device according to claim 1, characterized in that, The step of weightedly fusing multiple predicted values ​​based on the probability labels to obtain the weighted regression predicted value of the key pressure parameter at the current moment includes: The probability labels are normalized to obtain the normalized weights corresponding to each working condition category. Based on the predicted values ​​output by each regressor and the normalized weights corresponding to each operating condition category, the weighted regression predicted values ​​of the corresponding key pressure parameters at the current moment are determined.

7. The method for monitoring liquid sulfur blockage in a sulfur recovery device according to claim 1, characterized in that, The step of determining the deviation of the corresponding key pressure parameter based on the weighted regression predicted value and the actual value of the key pressure parameter includes: The deviation is determined based on the absolute error between the weighted regression prediction value of the corresponding key pressure parameter at the current moment and the actual value of the corresponding key pressure parameter.

8. A liquid sulfur blockage monitoring device for a sulfur recovery unit, characterized in that, include: The feature extraction unit is used to acquire real-time data for each operating condition category in the sulfur recovery unit at the current moment; the real-time data includes process parameters and key pressure parameters; utilizing... A pre-constructed time series large model is used to obtain the time series characteristics of the process parameters; wherein, the process parameters are input parameters reflecting the operating status of the sulfur recovery unit, including at least one of acid gas flow rate, mixed gas flow rate, combustion air flow rate, temperature and pressure parameters; the key pressure parameters are output parameters used to directly monitor the liquid sulfur blockage status, including at least one of main furnace pressure, first-stage Claus outlet pressure and second-stage Claus outlet pressure. The operating condition identification unit is used to fuse the process parameters and the time-series features, and input the fused features into a pre-trained time-series clusterer to obtain probability labels representing the likelihood that the fused features belong to each operating condition category; the fused features are comprehensive feature vectors formed by concatenating the process parameters and the time-series features, and simultaneously include static process attributes and dynamic time-series patterns; the time-series clusterer is a clusterer trained based on an unsupervised learning model, used to classify the fused features according to operating conditions; The weighted prediction unit inputs the process parameters in parallel to multiple pre-trained regressors to obtain multiple predicted values ​​of the key pressure parameters. Based on the probability labels, it performs weighted fusion of the multiple predicted values ​​to obtain the weighted regression prediction value of the key pressure parameters at the current time. Each regressor corresponds to a working condition category, and each regressor outputs a set of predicted values ​​of key pressure parameters. The anomaly detection unit determines that liquid sulfur blockage has occurred if the deviation of any key pressure parameter exceeds its corresponding preset deviation threshold, based on the deviation between the weighted regression predicted value and the actual value of the key pressure parameter.

9. The liquid sulfur blockage monitoring device for the sulfur recovery unit according to claim 8, characterized in that, The process parameters include at least one of the following parameters: Acid gas flow rate, mixed gas flow rate, combustion air flow rate, combustion air temperature, combustion air pressure, natural gas flow rate, natural gas flow rate and temperature, natural gas flow rate and pressure, fuel gas flow rate, fuel gas temperature, and fuel gas pressure; The key pressure parameter includes at least one of the following parameters: Main furnace pressure, first-stage Claus outlet pressure, and second-stage Claus outlet pressure.

10. The liquid sulfur blockage monitoring device for the sulfur recovery unit according to claim 8, characterized in that, It also includes an offline modeling unit, which is used for: The missing values ​​in the historical normal operating data of the sulfur recovery unit are filled in layers to obtain the filled historical data. The outliers in the filled historical data are then intelligently corrected to obtain the preprocessed historical data. An automated machine learning framework is used to select the optimal model from multiple time series models and train it to construct the large time series model for extracting the time series features. Using the trained time series large model, time series features are extracted from the preprocessed historical data, and the extracted time series features are fused with the preprocessed historical data to form historical fusion features; Unsupervised clustering is performed on the historical fusion features to divide them into multiple working condition categories in order to complete the training of the temporal clusterer; Based on the operating condition categories divided by the time-series clusterer of the preprocessed historical data, a regressor is trained for each operating condition category; wherein, the regressor adopts a multilayer perceptron structure and is used to predict key pressure parameters based on the input process parameters. Based on the prediction errors of multiple regressors on the historical normal operating data, the preset deviation thresholds corresponding to different key pressure parameters are determined by the 3σ rule.

Images