A method and system for diagnosing plugging and deactivation of an scr catalyst

Abstract

The application discloses a diagnosis method and system for SCR catalyst blockage and deactivation, belongs to the technical field of denitration catalysts, and solves the problem of how to improve the accuracy and response speed of SCR catalyst blockage and activity diagnosis. x The application extracts dynamic characteristics of multi-dimensional key parameters as a multi-dimensional feature vector of a diagnosis model input, combines residual analysis models through fusion analysis of multi-dimensional sensing data and LSTM prediction, early identifies, trend judges and classifies diagnoses of possible blockage or deactivation of the catalyst, and generates corresponding maintenance suggestions, thereby improving the accuracy and response speed of SCR catalyst blockage and activity diagnosis, and being suitable for intelligent operation and maintenance management of an environmental protection island SCR system in multiple industries such as thermal power, steel and cement.

Description

Technical Field

[0001] This invention belongs to the field of denitrification catalysis technology, and relates to a diagnostic method and system for SCR catalyst blockage and deactivation. Background Technology

[0002] Selective catalytic reduction (SCR) technology is a widely used flue gas denitrification process in industries such as thermal power, steel, and cement. Its core principle is to use a catalyst at a certain temperature to react a reducing agent (such as ammonia or urea) with nitrogen oxides (NOx) in the flue gas. x A reduction reaction occurs, producing harmless nitrogen gas and water, thus achieving NO reduction. x Highly efficient removal of nitrogen oxides. In SCR systems, the catalyst is a key component for achieving the reaction, and its performance directly determines the denitrification efficiency and emission levels. Currently, companies generally use regular manual inspections or shutdown maintenance to determine the catalyst condition, and some companies also use differential pressure or NOx control methods. x Over-limit alarms assist in judgment.

[0003] Although traditional monitoring and maintenance methods can maintain the operation of SCR systems to a certain extent, they still have the following obvious shortcomings in practical applications: (1) The monitoring dimensions are singular and lack comprehensive diagnostic capabilities. Current systems mostly use a single parameter as the basis for judgment, such as an increase in differential pressure or outlet NO. x (1) Exceeding the standard, easily affected by operational fluctuations and environmental factors, with high false alarm and false alarm rates, making it difficult to accurately judge the true state of catalyst blockage or deactivation. (2) Delayed state judgment and poor early warning capability. Most diagnoses are based on static thresholds or manual experience analysis, lacking dynamic trend tracking and prediction capabilities. They are often only detected after the catalyst performance has significantly declined, making timely intervention difficult and affecting the stable operation of the system. (3) Unable to quantify the degree of blockage or deactivation and lacking maintenance recommendations. Traditional methods cannot effectively identify the specific development degree of blockage and deactivation, nor can they provide clear operation and maintenance guidance, such as whether backflushing cleaning, module switching, or catalyst replacement is required. (4) Lack of remote and intelligent management capabilities. With the increasing environmental regulatory requirements and the advancement of enterprise intelligent upgrading, the original manual diagnostic methods are difficult to meet the modern operational needs of "predictive maintenance" and "data-driven decision-making", limiting the improvement of the intelligent level of the SCR system.

[0004] Existing technologies, such as the invention patent with publication number CN105893768, disclose a method for estimating the catalyst activity in a coal-fired boiler denitrification device. This method utilizes CFD software to simulate the internal flow field velocity and concentration distribution characteristics of the denitrification catalyst device, employs a conventional network sampling method for timed sampling of flue gas through measuring holes, processes the raw sampled data, and combines it with periodic catalyst testing data, process and structural parameters, to estimate the local activity of the catalyst. The obtained catalyst activity characteristics are then compared with a catalyst lifetime reference curve. However, while the above technology collects multiple types of parameters, it fails to capture the dynamic correlation between these parameters, and it is difficult to predict the trend evolution of catalyst activity in a timely manner based solely on periodic testing data. Summary of the Invention

[0005] The technical problem to be solved by this invention is how to improve the accuracy and response speed of SCR catalyst blockage and activity diagnosis.

[0006] The present invention solves the above-mentioned technical problems through the following technical solutions:

[0007] A diagnostic method for SCR catalyst blockage and deactivation includes the following steps:

[0008] S1, real-time acquisition of multi-dimensional key parameters of the SCR system;

[0009] S2, preprocessing the collected multidimensional key parameters;

[0010] S3: Extract dynamic features of multidimensional key parameters, perform weighted processing and alignment to form a multidimensional feature vector;

[0011] S4, construct and train the LSTM model, input the multi-dimensional feature vector, and output the predicted value of the multi-dimensional feature vector at the next time step;

[0012] S5, calculate the residual vector between the input and output, and generate anomaly scores based on weighted Euclidean distance;

[0013] S6. Based on the abnormal scoring results and residual vector weight allocation rules, diagnose the fault type and fault level, and verify the confidence level of the diagnosis results.

[0014] S7 outputs diagnostic results and corresponding maintenance recommendations.

[0015] Furthermore, the multidimensional key parameters mentioned in S1 include the SCR system inlet and outlet NO. x Concentration, catalyst bed inlet and outlet pressure, ammonia slip concentration, SCR system operating temperature, and flue gas velocity.

[0016] Furthermore, the dynamic features for extracting multidimensional key parameters described in S3 include:

[0017] S31, Extract the rate of change of pressure difference based on the trend of pressure difference change in the catalyst bed;

[0018] S32, according to NO x The export concentration shows a long-term upward trend, and the extraction shows an increasing trend.

[0019] S33, using the coefficient of variation to quantify the fluctuation characteristics of ammonia escape concentration, and extracting the coefficient of variation of ammonia escape;

[0020] S34, extracting catalytic reaction efficiency indicators;

[0021] S35, weighting and fusing the dynamic characteristics of multidimensional key parameters;

[0022] S36, set threshold judgment conditions for the dynamic features of multidimensional key parameters, and perform feature threshold judgment.

[0023] Furthermore, in step S33, the mean and standard deviation of the ammonia slip concentration are calculated hourly, and the coefficient of variation is calculated by the ratio of the mean to the standard deviation of the ammonia slip concentration. This is represented by the ammonia slip concentration fluctuation characteristics, and the ammonia slip concentration fluctuation rate is calculated based on the ratio of the maximum and minimum differences in ammonia slip concentration. This can be represented using the following logic:

[0024]

[0025]

[0026] In the formula, The coefficient of variation represents the ammonia escape concentration. This represents the standard deviation of ammonia slip concentration calculated from the time series. This represents the average ammonia slip concentration calculated from the time series. Indicates the ammonia escape concentration. Maximum ammonia escape concentration This represents the minimum ammonia escape concentration. This indicates the fluctuation rate of ammonia escape concentration.

[0027] Further, S35 includes the following steps:

[0028] S351, based on historical operating data and known system state labels, calculate the correlation score between each feature and the target state. ;

[0029] S352, normalize the original relevance scores to obtain the final weight coefficients. ;

[0030] S353, at each sampling time The feature vectors are weighted according to the weight coefficients to obtain the weighted feature vectors. This can be represented using the following logic:

[0031]

[0032] in, for The weighted feature vectors at each time step, For the first The eigenvectors at time... The value of , For the first Normalized weight coefficients of each eigenvector. Indicates the number of eigenvectors;

[0033] S354 performs statistical and alignment processing on the weighted feature vectors according to a unified time window, specifically as follows:

[0034] Set a fixed time window length The sampled data is segmented using the length of the time window as the sliding step. For the weighted feature data within each time window, the mean, variance, and moving average within that window are calculated sequentially to obtain the time window feature vector. :

[0035]

[0036] in, This represents the moving average operation. Indicates taking The mean, Indicates taking variance Indicates taking The moving average;

[0037] The time window feature vectors of each feature are arranged sequentially in chronological order to form a time-series multidimensional weighted fusion feature matrix. :

[0038]

[0039] in, For multidimensional weighted fusion feature matrix, The number of time windows to look back. express The length of the time window.

[0040] Furthermore, the LSTM model described in S4 includes an input layer, an LSTM layer, a fully connected layer, and an output layer;

[0041] The input layer is used to receive multidimensional feature matrices. ,in T represents the length of the time window. The dimension representing the multidimensional weighted fusion feature;

[0042] The LSTM layer adopts a two-layer stacked structure, with 64 neurons in each layer and the activation function is tanh.

[0043] The fully connected layer is used to map the final hidden state of the LSTM to a dimension of . The output space has the same dimensions as the input multidimensional feature vector;

[0044] The output layer uses a linear activation function to generate the predicted value for the next time step. ;

[0045] The training of the LSTM model specifically involves: using historical data from the SCR system under normal operating conditions as training data, employing Huber loss as the loss function, selecting the AdamW optimizer to optimize model parameters, validating model performance during training based on an early stopping strategy, suppressing overfitting through Dropout regularization, and representing the Huber loss using the following logic:

[0046]

[0047] In the formula, Indicates the loss value. Indicates the threshold parameter. express The multidimensional feature vectors observed at any given moment. express The predicted feature vector at time step.

[0048] Further, S5 includes the following steps:

[0049] S51, calculate based on the trained LSTM model Predicted value at time Compared with actual observation input The residual vector between This can be represented using the following logic:

[0050]

[0051] In the formula, for The multidimensional feature vectors observed at any given moment. , For the corresponding predicted feature vector, Let be the residual vector of the corresponding dimension, denoted as ;

[0052] S52, assign weights based on parameter importance. For the residual vector The elements of each dimension are weighted and anomaly scores are calculated. This can be represented using the following logic:

[0053]

[0054] In the formula, For the first 3D residual vector ;

[0055] S53, Adaptive dynamic threshold based on historical normal operating condition residual distribution To determine abnormal states, specifically: calculate the current residual score in real time. , and dynamic threshold When comparing, ≤ When the SCR system is in normal condition, it is determined that the system is in a normal state; when > When an anomaly is detected in the SCR system, step S6 is executed to further subdivide the anomaly level and assist in maintenance decision-making; wherein, the adaptive dynamic threshold... This represents the 95th percentile of the historical normal residual scores.

[0056] Further, S6 includes the following steps:

[0057] S61. Based on the contribution of different parameters in the residual vector, the fault type is distinguished using the following logical representation:

[0058] S611, when the pressure difference change rate in S52 When the residual weight is >60% and the residual of the catalytic reaction efficiency η increases significantly, catalyst blockage is diagnosed.

[0059] S612, when the ammonia escape variation coefficient in S52 When the residual weight is >50% and the catalytic reaction efficiency η continues to decrease, catalyst deactivation is diagnosed.

[0060] S613, when S611 and S612 are satisfied at the same time, the catalyst is diagnosed to have both blockage and deactivation.

[0061] S614, when neither S611 nor S612 is satisfied, the catalyst is diagnosed as normal;

[0062] S62, quantify the fault level based on the anomaly scoring results, using the following logical representation:

[0063] when And duration At that time, the catalyst state is determined to be either initial blockage or initial deactivation;

[0064] when or and At that time, the catalyst condition is judged to be severely deactivated or severely blocked;

[0065] Here, "Score" refers to the anomaly score calculated by the model based on the input feature vector at the current moment. This represents the initial threshold, calculated as the 90th percentile of the residual score from historical normal data. The severity threshold is represented by the 70th percentile of the residual score from historical fault data. Indicates the duration threshold. This indicates that the abnormal score has reached its first peak. The duration of the fault state;

[0066] S63. Perform confidence verification on the diagnostic results and improve the reliability of the diagnostic results by introducing multi-indicator cross-validation.

[0067] Furthermore, S7 specifically includes:

[0068] The current catalyst status is output based on the model diagnostic results, including normal, initial blockage, severe blockage, and catalyst deactivation.

[0069] When the output catalyst is in normal condition, no maintenance is recommended; when the output catalyst is in initial blockage condition, it is recommended to optimize the soot blowing frequency and check the soot blower; when the output catalyst is in severe blockage condition, it is recommended to shut down and flush with high pressure or switch to a backup module; when the output catalyst is in initial deactivation or severe deactivation condition, it is recommended to calibrate the ammonia injection control, evaluate regeneration, or replace it.

[0070] This invention also provides a diagnostic system for SCR catalyst blockage and deactivation, comprising:

[0071] The data acquisition module is used to collect multi-dimensional key parameters of the SCR system in real time;

[0072] The data preprocessing module is used to preprocess the collected multidimensional key parameters;

[0073] The feature extraction module is used to extract dynamic features of multidimensional key parameters, perform weighted processing and alignment, and form a multidimensional feature vector.

[0074] The LSTM prediction module is used to build and train an LSTM model, taking a multi-dimensional feature vector as input and outputting the predicted value of the multi-dimensional feature vector at the next time step.

[0075] The residual analysis module is used to calculate the residual vectors between the input and output, and generate anomaly scores based on weighted Euclidean distance;

[0076] The status classification module is used to diagnose fault types and fault levels based on anomaly scoring results and residual vector weight allocation rules, and to verify the confidence level of the diagnostic results.

[0077] The results output module is used to output diagnostic results and corresponding maintenance suggestions.

[0078] The advantages of this invention are:

[0079] This invention comprehensively analyzes the trend changes in catalyst bed pressure difference and NO. x By extracting dynamic features of key operating indicators such as outlet concentration trend changes, ammonia slip fluctuation characteristics, and catalytic reaction efficiency, multidimensional key parameters are used as multidimensional feature vectors input to the diagnostic model. Through the fusion analysis of multidimensional sensor data and LSTM prediction combined with residual analysis models, early identification, trend judgment, and graded diagnosis of potential catalyst blockage or deactivation are achieved, and corresponding maintenance suggestions are generated. This enables intelligent identification and fault early warning of SCR catalyst operating status, improves the operational reliability and environmental compliance of denitrification systems, and also improves the accuracy and response speed of SCR catalyst blockage and activity diagnosis. It is applicable to the operation monitoring and intelligent maintenance of SCR systems in various large-scale stationary pollution sources and has broad engineering application value.

[0080] This invention integrates pressure difference and NO x By integrating multiple key indicators such as ammonia slip and multi-parameter diagnosis, the LSTM-based prediction model can predict and analyze the future system state, provide early warnings before catalyst blockage or deactivation seriously affects efficiency, reserve sufficient maintenance windows, and reduce false alarms and false negatives by using time series trend analysis and model discrimination methods, thus improving diagnostic accuracy.

[0081] This invention quantifies the degree of blockage or deactivation and, based on confidence level verification, uses knowledge from the field of flue gas denitrification to perform secondary verification of the LSTM model and the diagnostic results of residual analysis. This ensures that the model output not only conforms to the data pattern but also to physical and chemical laws, avoiding misclassification or prediction failure, and improving the reliability and accuracy of the diagnostic model in actual production environments.

[0082] This invention optimizes deployment by bringing the trained LSTM model to the actual production environment. Lightweighting the LSTM model ensures good computational efficiency even when running on edge devices, improving real-time hardware responsiveness. Simultaneously, the LSTM model is updated online, with new data periodically added to the training process to gradually update the model's parameters. The model can be fine-tuned periodically with new data, and the catalyst aging trend is updated in real time, providing more accurate predictions. The diagnostic model provided by this invention can be embedded in control systems such as DCS / PLC, featuring a clear structure, interpretable logic, and ease of maintenance, supporting rapid replication and application across multiple scenarios and units.

[0083] This invention can effectively prevent sudden efficiency drops or emission exceedances in SCR systems, helping users achieve stable compliance with environmental protection standards, while extending catalyst lifespan and reducing operation and maintenance costs. It is particularly suitable for intelligent operation and maintenance management of SCR systems in environmental protection islands in various industries such as thermal power, steel, and cement, and is a key link in promoting the development of environmental protection equipment towards self-diagnosis, self-maintenance, and intelligence. Attached Figure Description

[0084] Figure 1 This is a schematic diagram of the diagnostic model data flow according to Embodiment 1 of the present invention;

[0085] Figure 2 This is a schematic diagram of the diagnostic system for SCR catalyst blockage and deactivation according to Embodiment 1 of the present invention;

[0086] Figure 3 This is a schematic diagram of the engineering application deployment of Embodiment 1 of the present invention. Detailed Implementation

[0087] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0088] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments:

[0089] Example 1

[0090] like Figure 1 Specifically, a diagnostic method for SCR catalyst blockage and deactivation is disclosed, including:

[0091] S1 collects multi-dimensional key parameters of the SCR system in real time.

[0092] The multidimensional key parameters include SCR system inlet and outlet NO. x Concentration, catalyst bed inlet and outlet pressure, ammonia (NH3) slip concentration, SCR system operating temperature, flue gas velocity, and other auxiliary operating condition data. Among these, NO... x NO concentration is set at the inlet and outlet of the SCR system. x The inlet and outlet pressures of the catalyst bed are obtained through the ΔP differential pressure sensor, the ammonia slip concentration is obtained through the ammonia slip monitor, and the operating temperature and flue gas velocity of the SCR system are obtained through temperature and flow rate sensors, respectively. Other auxiliary operating condition data can be obtained through the DCS / PLC system or through the corresponding sensors built into the SCR system.

[0093] In addition, various sensors are connected through DCS systems, PLC controllers or edge computing devices to collect multi-dimensional key parameters during the operation of the SCR system in real time, ensuring the timeliness and accuracy of the data.

[0094] S2, preprocesses the collected multidimensional key parameters.

[0095] To improve the accuracy of subsequent multidimensional parameter fusion analysis, step S2 performs preprocessing operations on the collected raw multidimensional key data, including outlier removal and repair, missing value interpolation and imputation, data smoothing and denoising, and standardization and normalization. In this embodiment, the above data preprocessing operations can be selectively performed according to the actual data quality and application requirements.

[0096] In this embodiment, existing data preprocessing methods can be used to preprocess multidimensional key data. For example, in the outlier removal and repair stage, the boxplot method can be used to detect outliers using quartiles (IQR); in the missing value interpolation and imputation stage, linear interpolation can be used to interpolate linearly from adjacent known data points, or sliding window regression can be used to imput missing values ​​through a local regression model; in the data smoothing and denoising stage, wavelet transform or Kalman filtering can be used to eliminate noise in the dynamic system.

[0097] S3 extracts the dynamic features of multidimensional key parameters, performs weighted processing and alignment, and forms a multidimensional feature vector.

[0098] In step S3, the extraction of dynamic features of multidimensional key parameters specifically involves:

[0099] S31, based on catalyst bed pressure difference The trend change was analyzed, and the rate of change of pressure difference was extracted. .

[0100] Catalyst bed resistance variation is a factor in determining blockage; this can be assessed by extracting the catalyst bed pressure differential change rate. This is used to monitor changes in catalyst bed resistance, thereby determining whether there are signs of blockage. Specifically, the catalyst bed inlet pressure is first collected in a time series. and export pressure And solve for the inlet and outlet pressure difference. Then calculate the pressure difference change between the current moment and the previous moment, using the following logic:

[0101]

[0102]

[0103] In the formula, In this embodiment, the sampling time interval is indicated. It can also represent the sampling period. express The pressure difference between the entrance and exit at any given moment.

[0104] S32, according to NO x Export concentration shows a long-term upward trend, and extraction shows an increasing trend.

[0105] The growth trend This method is used to identify whether catalytic efficiency has decreased or whether reactant is sufficient. Specifically, this embodiment uses a sliding window averaging method, with a sliding average of 5 minutes or 10 minutes as the window to smooth out noise interference and identify NO. x The export concentration continues to rise, and the NO concentration at the current moment is compared with that at the previous moment. x The average export concentration is expressed using the following logic:

[0106]

[0107] In the formula, Indicates the current Exit of Time NO x The average value of the concentration over a sliding window. This indicates that in the calculation of the window moving average, Exit of Time NO x Concentration, where N represents the size of the sliding window. This indicates the current time compared to a certain period of time in the past (in this embodiment, it refers to...). Time and The difference in NOx moving average concentration compared to time T before is used to identify an upward trend in concentration. express Exit NO T time before time T x Concentration, where T represents the time span for trend detection.

[0108] S33, using the coefficient of variation to quantify the fluctuation characteristics of ammonia escape concentration, extracting the coefficient of variation of ammonia escape. .

[0109] This embodiment introduces the coefficient of variation (CV) and volatility of ammonia slip concentration as characteristic parameters for the first time, to characterize the control stability of the ammonia injection system and the changes in catalyst conversion capacity. Compared with the traditional method of using a fixed concentration threshold, this method is more sensitive to early performance degradation trends and has advantages such as strong adaptability, high accuracy in anomaly identification, and low engineering deployment threshold, which can significantly improve the intelligent diagnostic accuracy and response time of the SCR system.

[0110] In this embodiment, the ammonia slip concentration fluctuation characteristics are used to identify problems such as decreased catalyst activity and abnormal ammonia-nitrogen ratio. Specifically, the mean and standard deviation of the ammonia slip concentration are calculated hourly, and the coefficient of variation is calculated by the ratio of the mean to the standard deviation of the ammonia slip concentration. This represents the fluctuation characteristics of ammonia escape concentration. The ammonia escape concentration volatility is then calculated based on the ratio of the maximum to minimum difference in ammonia escape concentration. This can be represented using the following logic:

[0111]

[0112]

[0113] In the formula, The coefficient of variation represents the ammonia escape concentration. This represents the standard deviation of ammonia slip concentration calculated from the time series. This represents the average ammonia slip concentration calculated from the time series. Indicates the ammonia escape concentration. Maximum ammonia escape concentration This represents the minimum ammonia escape concentration. This indicates the fluctuation rate of ammonia escape concentration.

[0114] S34, Extracting catalytic reaction efficiency indicators .

[0115] The catalytic reaction efficiency index is used to measure the real-time performance of the SCR system in removing NOx. Specifically, the removal efficiency is calculated in real time by collecting the inlet and outlet NOx concentrations, and is expressed using the following logic:

[0116]

[0117] In the formula, Indicates the efficiency of the catalytic reaction. Indicates the inlet NOx concentration. This indicates the NOx concentration at the export site.

[0118] Furthermore, the dynamic characteristics of the multidimensional key parameters extracted in steps S31-S34 (including the rate of change of pressure difference in the catalyst bed) NOx export concentration growth trend Ammonia escape concentration fluctuation characteristics and Based on features such as catalytic reaction efficiency η, this embodiment uses feature importance analysis, weight allocation, time alignment, and fusion to form a multidimensional weighted fusion feature matrix for model diagnosis. The weighting and alignment process specifically includes the following steps:

[0119] S35 involves weighting and fusing the dynamic characteristics of multidimensional key parameters. Specifically, S35 includes the following steps:

[0120] S351, Feature Importance Calculation: Specifically, based on historical operating data and known system state labels (blockage, inactivation, ammonia injection anomaly, etc.), the correlation score between each feature and the target state is calculated. .

[0121] In this embodiment, the target state is used to monitor the system operation results of feature importance calculation, or the labeled fault category label; the target state includes, but is not limited to, known fault types or normal states such as catalyst blockage, catalyst deactivation, and ammonia injection abnormality. These states are derived from historical operation data and can guide the correlation analysis between features and states through manual annotation, expert diagnosis, or automatic system identification results.

[0122] In this embodiment, the correlation score The feature importance score can be calculated using Pearson correlation coefficient, mutual information coefficient, or any machine learning model such as random forest or gradient boosting tree, and is denoted as . ,in, For the first 1 eigenvector This represents the state label of the corresponding feature vector. This is the original relevance score.

[0123] S352, Weight Normalization, specifically involves normalizing the original correlation scores to obtain the final weight coefficients. This can be represented using the following logic:

[0124]

[0125] in, For the first Normalized weight coefficients of each eigenvector. Indicates the first The original relevance scores corresponding to each feature vector. This represents the number of feature vectors, and the weights involved later in this embodiment. All indicate the first The weight coefficients of the eigenvectors are normalized and satisfy the constraints. .

[0126] S353, weighted processing, specifically at each sampling time. The feature vectors are weighted according to the weight coefficients to obtain the weighted feature vectors. This can be represented using the following logic:

[0127]

[0128] in, for The weighted feature vectors at each time step, For the first The eigenvectors at time... The value of .

[0129] S354, Time Window Alignment and Blending.

[0130] In this embodiment, to ensure the consistency of different features over time, the weighted feature vectors are statistically analyzed and aligned according to a unified time window.

[0131] Specifically, a fixed time window length is set. (Unit: minutes), and the sampled data is segmented using the length of this time window as the sliding step. For the weighted feature data within each time window, the mean, variance, and moving average within that window are calculated sequentially to obtain the time window feature vector. :

[0132]

[0133] in, This represents the moving average operation. Indicates taking The mean, Indicates taking variance Indicates taking The moving average.

[0134] The time window feature vectors of each feature are arranged sequentially in chronological order to form a time-series multidimensional weighted fusion feature matrix. :

[0135]

[0136] in, For multidimensional weighted fusion feature matrix, The number of time windows to look back. express The length of the time window.

[0137] The multidimensional weighted fusion feature matrix As the input sequence for the LSTM diagnostic model, it achieves complete alignment and synchronous update of different features in the time dimension, ensuring the consistency and comparability of the model input data.

[0138] S36, set threshold judgment conditions for the dynamic features of multidimensional key parameters, and perform feature threshold judgment.

[0139] In this embodiment, threshold judgment conditions are set for the multidimensional key features extracted in steps S31 to S34 to assist in the intelligent diagnosis of catalyst blockage and deactivation. Specifically, this includes the following:

[0140] S361, when the coefficient of variation of ammonia escape concentration ( ) or volatility ( When the value exceeds 1.5 to 2 times the historical average, it is judged that the ammonia injection system is fluctuating abnormally or the catalyst activity is decreasing.

[0141] S362, when the catalytic reaction efficiency η continues to decrease or falls below the set threshold (e.g., η<80%), it indicates catalyst deactivation, insufficient ammonia injection, or abnormal temperature.

[0142] S363, When the moving average of NOx export concentration continues to increase and the increase exceeds the threshold. If the catalytic efficiency is degraded or the ammonia injection is insufficient, it is determined that the catalytic efficiency has degrade

[0143] S364, when the catalyst bed pressure difference change rate Exceeding the threshold If the duration is ≥ T, a blockage trend is identified.

[0144] S365: When any of the threshold judgment conditions in S361 to S364 are met, the system determines that there is a potential fault such as abnormal fluctuation of the ammonia injection system or decreased catalyst activity. Step S4 is executed, and the multi-dimensional feature vector of the corresponding time period is sent to the subsequent diagnostic model for in-depth analysis and fault classification. If none of the threshold judgment conditions in S361 to S364 are triggered, the system is considered to be in a stable range. S4 does not need to be executed, and no data needs to be submitted to the diagnostic model, so as to reduce unnecessary calculations and false alarms.

[0145] In this embodiment, the threshold judgment conditions provided in steps S361 to S364 are used for pre-screening of the intelligent diagnostic model. Through the two-level judgment mechanism of threshold screening followed by in-depth diagnosis, redundant calculations of the model under completely normal working conditions can be avoided, while ensuring that the intelligent diagnostic process can be entered in time when early abnormal signs appear, thereby improving the overall accuracy and response speed of the diagnosis.

[0146] S4. Construct and train an LSTM model, input a multi-dimensional feature vector, and output the predicted value of the multi-dimensional feature vector at the next time step.

[0147] In the scenario of intelligent diagnosis of SCR catalyst blockage, the "blockage" or "deactivation" state of the target is identified from multi-dimensional time-series data (such as differential pressure, NOx concentration, ammonia slip, etc.). Taking into account factors such as the real-time performance, robustness, deployment difficulty and interpretability of the state diagnosis, this invention uses LSTM prediction combined with residual analysis as the diagnostic model. By leveraging the characteristic of LSTM to support multi-dimensional time-series input, the spatiotemporal correlation between multi-dimensional key parameters is learned through a gating mechanism, and the multi-dimensional feature vector is directly used as the data for each time step of the LSTM input layer.

[0148] Specifically, the LSTM model includes an input layer, an LSTM layer, a fully connected layer, and an output layer.

[0149] The input layer is used to receive the multidimensional feature matrix output by S3. ,in T represents the length of the time window (T = 24 hours). The dimension representing the multidimensional weighted fusion feature. Multidimensional feature matrix The input data sequence for the LSTM model contains multi-dimensional statistical information of historical time series features, covering the rate of change of pressure difference. NOx concentration trend ammonia escape coefficient of variation Catalytic reaction efficiency Several key features of weighted fusion, etc.

[0150] The LSTM layer adopts a two-layer stacked structure with 64 neurons in each layer and the activation function is tanh, so as to fully capture the long-term dependencies and dynamic trends of the input features.

[0151] The fully connected layer is used to map the final hidden state of the LSTM to a dimension of . The output space has the same dimensions as the input multidimensional feature vector;

[0152] The output layer uses a linear activation function (i.e., no activation function) to generate the predicted value for the next time step. This is to meet the needs of continuous value prediction.

[0153] Furthermore, the constructed LSTM model is trained. Specifically, historical data under normal operating conditions of the SCR system is used as training data, Huber loss is used as the loss function, AdamW optimizer is selected to optimize model parameters, early stopping strategy is used to verify model performance during training, and Dropout regularization algorithm is used to suppress overfitting.

[0154] In this embodiment, the Huber loss is represented using the following logic:

[0155]

[0156] In the formula, Indicates the loss value. Indicates the threshold parameter. express The multidimensional feature vectors observed at real time are used to balance the advantages of MSE and MAE by utilizing Huber loss, thereby improving the robustness of the model.

[0157] In this embodiment, the training data should cover historical data of the SCR system under normal operating conditions under different loads and seasons; the learning rate of the AdamW optimizer is... The weight decays to The early stopping strategy specifically involves terminating training early when the validation loss fails to decrease for 10 consecutive rounds, indicating that the model performance will no longer improve. This prevents overfitting of the model to the training data and saves computational resources. The Dropout regularization has a Dropout probability of 0.2, meaning that each neuron has a 20% probability of being dropped. During training, some neurons in the network are temporarily disabled and do not participate in the forward and backward propagation of the current round.

[0158] S5 calculates the residual vectors of the input and output, and generates anomaly scores based on weighted Euclidean distance.

[0159] In this embodiment, a two-layer LSTM model is used to predict the multidimensional feature vectors of the SCR system. The prediction results are used to calculate the residual vector between the actual observed data and the predicted values, thereby achieving anomaly detection. This avoids the labeling dependency and model overfitting problems caused by direct classification, and improves the model's generalization ability and its ability to identify unknown anomalies. Specifically, step S5 includes the following steps:

[0160] S51, Calculate the residual vector Specifically, it involves calculating based on the trained LSTM model. Predicted value at time Compared with actual observation input The residual vector between This can be represented using the following logic:

[0161]

[0162] In the formula, for The multidimensional feature vectors observed at any given moment. , The corresponding predicted feature vector is output by the trained LSTM model. Let be the residual vector of the corresponding dimension, denoted as .

[0163] S52, residual vector weighting and anomaly score calculation, specifically using weighted Euclidean distance to enhance the rate of change of pressure difference. The sensitivity of key characteristics such as catalytic reaction efficiency η to anomalies is determined by assigning weights based on parameter importance. For the residual vector The elements of each dimension are weighted and anomaly scores are calculated. This can be represented using the following logic:

[0164]

[0165] In the formula, For the first 3D residual vector .

[0166] S53, Dynamic Threshold Setting and Anomaly Detection: To adapt to the time-varying characteristics of SCR system operation and environmental fluctuations, an adaptive dynamic threshold based on the residual distribution of historical normal operating conditions is adopted. This is used to determine abnormal states and avoid fixed thresholds being affected by fluctuations in operating conditions.

[0167] In this embodiment, the dynamic threshold Setting the threshold to the 95th percentile of historical normal residual scores allows for automatic adjustment of detection sensitivity, reducing false alarm rates caused by fluctuations in operating conditions; the determination of abnormal states specifically refers to:

[0168] Real-time calculation of current residual score , and dynamic threshold When comparing, ≤ When the SCR system is in normal condition, it is determined that the system is in a normal state; when > When an anomaly is detected in the SCR system, the anomaly level is further subdivided based on the primary and critical thresholds preset in S6 to assist in maintenance decision-making. The dynamic threshold mechanism proposed in this embodiment effectively overcomes the problem of fixed thresholds being sensitive to fluctuations in operating conditions, leading to false alarms or missed alarms, and improves the stability and robustness of the diagnostic system.

[0169] Furthermore, based on deployment optimization, the trained LSTM model is deployed to actual production environments, including edge devices, industrial control systems, or resource-constrained equipment. This optimization adapts the model to industrial hardware by improving performance and reducing resource consumption. The deployment optimization not only includes lightweighting the LSTM model but also real-time execution of residual calculation and anomaly scoring, ensuring the overall diagnostic system operates efficiently and stably. Specifically, the deployment optimization includes:

[0170] Lightweighting of LSTM models involves compressing the number of model parameters and storage requirements through pruning and quantization.

[0171] In this embodiment, pruning specifically involves identifying parameters with minimal impact on performance based on the sparsity of certain weights, or evaluating the model before quantization to remove these less important parameters, thereby reducing the model size. Quantization specifically involves converting the weights and activation values, originally represented as floating-point numbers, into low-precision integers (e.g., INT8), reducing the number of bits required for representation, thus compressing the model's storage size and accelerating the computation process. Through pruning and quantization, the lightweight model can maintain good computational efficiency even when executed on low-precision hardware such as edge devices (e.g., industrial PLCs), improving the hardware's real-time response capabilities.

[0172] The LSTM model is updated online using incremental learning technology. The system periodically incorporates new data collected from the industrial field into the model training process, dynamically adjusting model parameters to reflect catalyst aging and changes in operating conditions in real time. This online update mechanism solves the problem of traditional offline-trained models struggling to adapt to long-term equipment operating changes, improving the accuracy and robustness of diagnostics.

[0173] In this embodiment, since catalysts age over time in industrial applications, traditional offline training cannot reflect this change in a timely manner. This embodiment uses online updates, allowing the model to be fine-tuned with new data periodically, updating the catalyst aging trend in real time and providing more accurate predictions.

[0174] S6, based on the anomaly scoring results and residual vector weight allocation rules, diagnoses the fault type and fault level, and verifies the confidence level of the diagnostic results. S6 includes the following steps:

[0175] S61, the fault type is distinguished based on the contribution of different parameters in the residual vector, using the following logical representation:

[0176] S611, when the pressure difference change rate in S52 When the residual weight is >60% and the residual of the catalytic reaction efficiency η increases significantly, catalyst blockage is diagnosed; wherein, the significant increase in the residual of the catalytic reaction efficiency η specifically means:

[0177] The residual value of catalytic reaction efficiency η When the value of the minimum absolute change ratio threshold and the statistical threshold is exceeded, the residual of the catalytic reaction efficiency η is considered to have increased significantly. In this embodiment, the minimum absolute change ratio threshold is taken as 0.05 for explanation, and is represented by the following logic:

[0178]

[0179] in, This is the historical normal average. The standard deviation threshold, Indicates the statistical threshold. and All data were calculated from historical data on the normal operation of the catalyst.

[0180] S612, when the ammonia escape variation coefficient in S52 When the residual weight is >50% and the catalytic reaction efficiency η continuously decreases, catalyst deactivation is diagnosed; specifically, the continuous decrease in catalytic reaction efficiency η is defined as follows: the catalytic reaction efficiency η exhibits a monotonically decreasing trend over M consecutive time steps (e.g., M=3 or 6), with a cumulative decrease of [missing value]. satisfy .

[0181] S613, when S611 and S612 are satisfied simultaneously, the catalyst is diagnosed to be both blocked and deactivated.

[0182] S614: When neither S611 nor S612 is satisfied, the catalyst is diagnosed as normal.

[0183] S62, quantify the fault level based on the anomaly scoring results, using the following logical representation:

[0184] when And duration (like When the catalyst is initially blocked or initially deactivated, its condition is determined.

[0185] when or and At that time, the catalyst condition is judged to be severely deactivated or severely blocked;

[0186] Here, "Score" refers to the anomaly score calculated by the model based on the input feature vector at the current moment. This represents the initial threshold, calculated as the 90th percentile of the residual score from historical normal data. The severity threshold is represented by the 70th percentile of the residual score from historical fault data. Indicates the duration threshold. This indicates that the abnormal score has reached its first peak. The duration of the fault state is as follows, as shown in Table 1:

[0187] Table 1 Catalyst Failure Levels

[0188]

[0189] S63, perform confidence verification on the diagnostic results. By introducing multi-indicator cross-validation, the reliability of the diagnostic results is improved, using the following logical representation:

[0190] When the diagnostic result indicates catalyst blockage, verify the differential pressure change rate. The negative correlation with flue gas velocity; if the differential pressure change rate increases significantly when the flue gas velocity is stable or decreases less than the set value, the diagnostic result is considered to be consistent with the actual fluid dynamics law and is judged to have high confidence.

[0191] When the diagnosis result is catalyst deactivation, verify the logical consistency between the ammonia-nitrogen molar ratio setpoint and the ammonia slip fluctuation; if the ammonia-nitrogen molar ratio setpoint remains stable and the ammonia slip concentration fluctuation rate increases significantly, then the diagnosis result is considered to be consistent with the chemical mechanism of the catalytic reaction and is judged to be of high confidence.

[0192] If the verification results are inconsistent with the diagnostic conclusion, they are marked as low confidence and a secondary diagnostic process is triggered. This process can call up a backup mechanism model or compare with similar historical operating conditions, and prompt manual review to avoid unnecessary maintenance operations due to misjudgment.

[0193] Based on confidence level verification, the diagnostic results of the LSTM model and residual analysis are validated a second time using knowledge from the field of flue gas denitrification. This ensures that the model output not only conforms to the data pattern but also to physical and chemical laws, avoiding misclassification or prediction failure. This improves the reliability and accuracy of the diagnostic model in actual production environments, reduces false alarms and missed alarms caused by operating condition fluctuations or sensor noise, avoids diagnostic conclusions that are physically invalid, enhances the availability and safety of automated control systems, reduces economic losses and safety risks caused by misjudgments, and increases the acceptability of diagnostic results in industry expert reviews or safety regulatory audits.

[0194] S7 outputs diagnostic results and corresponding maintenance recommendations.

[0195] Based on the model's diagnostic results, the current catalyst status is output, including normal, initial blockage, severe blockage, and catalyst deactivation, and corresponding maintenance suggestions are provided as shown in Table 2 below:

[0196] Table 2 Catalyst Condition Diagnosis Results and Maintenance Recommendations

[0197]

[0198] As shown in Table 2, when the output catalyst is in normal condition, no maintenance is required; when the output catalyst is in initial blockage condition, it is recommended to optimize the soot blowing frequency and check the soot blower; when the output catalyst is in severe blockage condition, it is recommended to shut down the unit for high-pressure flushing or switch to a backup module; when the output catalyst is in deactivated condition (including initial deactivation and severe deactivation), it is recommended to calibrate the ammonia injection control, evaluate regeneration, or replace the catalyst.

[0199] This invention comprehensively analyzes the trend changes in catalyst bed pressure difference and NO. x By extracting dynamic features of key operating indicators such as outlet concentration trend changes, ammonia slip fluctuation characteristics, and catalytic reaction efficiency, multidimensional key parameters are used as multidimensional feature vectors input to the diagnostic model. Through the fusion analysis of multidimensional sensor data and LSTM prediction combined with residual analysis models, early identification, trend judgment, and graded diagnosis of potential catalyst blockage or deactivation are achieved, and corresponding maintenance suggestions are generated. This enables intelligent identification and fault early warning of SCR catalyst operating status, improves the operational reliability and environmental compliance of denitrification systems, and enhances the accuracy and response speed of SCR catalyst blockage and activity diagnosis. It is applicable to the operation monitoring and intelligent maintenance of SCR systems in various large-scale stationary pollution sources and has broad engineering application value. It is particularly suitable for the intelligent operation and maintenance management of SCR systems in environmental protection islands in multiple industries such as thermal power, steel, and cement. It is a key link in promoting the development of environmental protection equipment towards self-diagnosis, self-maintenance, and intelligence.

[0200] As attached Figure 3 As shown, the present invention provides an engineering deployment scheme for an intelligent diagnostic system that adopts a layered architecture to achieve multi-level and multi-node collaborative work, including modules such as a mobile APP, a web monitoring screen, an operation and maintenance center, a cloud diagnostic platform, plant-level monitoring servers, edge computers, and unit SCR systems.

[0201] The specific deployment structure is as follows:

[0202] The mobile app and web-based monitoring dashboard provide maintenance personnel with real-time data access, alarm information display, and remote control functions, improving monitoring convenience and response speed.

[0203] As a centralized management and dispatch center, the operations and maintenance center gathers data and alarm information from various plants for unified management.

[0204] The cloud diagnostic platform is responsible for intelligent diagnostics of multi-dimensional time-series data collected from multiple plants, including residual calculation, anomaly scoring and fault warning based on a lightweight two-layer LSTM model, and supports online model updates and dynamic threshold adjustments.

[0205] Each plant has its own plant-level monitoring server, which is responsible for collecting real-time operational data of the plant and providing processing interfaces to edge computers.

[0206] Edge computers perform tasks such as local data preprocessing, running lightweight LSTM models, residual calculation, and anomaly detection, enabling low-latency local diagnostic capabilities.

[0207] The unit's SCR system is responsible for the specific control and execution of the denitrification process.

[0208] The central control center enables centralized control of the operating status of each plant through remote monitoring, and supports unified management of multiple units across regions.

[0209] This layered architecture not only ensures the real-time nature of data acquisition and the accuracy of diagnostics, but also fully considers the hardware resource limitations and network bandwidth fluctuations in industrial settings. By deploying lightweight models and residual calculations on edge computing nodes, it achieves localized and real-time intelligent diagnostics, effectively reducing reliance on cloud platform computing resources. Simultaneously, it supports an online model update mechanism to dynamically adapt to catalyst aging and changes in the operating environment.

[0210] In summary, Appendix Figure 3 The deployment scheme shown fully demonstrates the application value and engineering feasibility of the intelligent diagnostic system of this invention in a real industrial environment, and meets the needs of collaborative monitoring and fault early warning for multiple units and multiple factories.

[0211] The present invention also provides a diagnostic system applied to the above-mentioned SCR catalyst blockage and deactivation diagnostic method, comprising:

[0212] The data acquisition module is used to collect multi-dimensional key parameters of the SCR system in real time.

[0213] The data preprocessing module is used to preprocess the collected multidimensional key parameters.

[0214] The feature extraction module is used to extract dynamic features of multidimensional key parameters, perform weighted processing and alignment, and form a multidimensional feature vector.

[0215] The LSTM prediction module is used to build and train an LSTM model, taking a multi-dimensional feature vector as input and outputting the predicted value of the multi-dimensional feature vector at the next time step.

[0216] The residual analysis module is used to calculate the residual vectors between the input and output, and generate anomaly scores based on weighted Euclidean distance.

[0217] The status classification module is used to diagnose fault types and fault levels based on anomaly scoring results and residual vector weight allocation rules, and to verify the confidence level of the diagnostic results.

[0218] The results output module is used to output diagnostic results and corresponding maintenance suggestions.

[0219] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for diagnosing SCR catalyst blockage and deactivation, characterized in that, Includes the following steps: S1, real-time acquisition of multi-dimensional key parameters of the SCR system; S2, preprocessing the collected multidimensional key parameters; S3, extract the dynamic features of multidimensional key parameters, perform weighted processing and alignment to form a multidimensional feature vector; wherein, the weighted processing and alignment to form a multidimensional feature vector includes: weighting and fusing the dynamic features of the multidimensional key parameters, specifically: Based on historical operational data and known system state labels, the correlation score between each feature and the target state is calculated. ; The original relevance scores are normalized to obtain the final weight coefficients. ; Weighted processing, specifically at each sampling time The feature vectors are weighted according to the weight coefficients to obtain the weighted feature vectors. ; Time window alignment and merging; S4, construct and train the LSTM model, input the multi-dimensional feature vector, and output the predicted value of the multi-dimensional feature vector at the next time step; S5 calculates the residual vectors between the input and output, and generates anomaly scores based on weighted Euclidean distance, specifically: Calculate the residual vector We employ weighted Euclidean distance to enhance the sensitivity of key features to anomalies, specifically by assigning weights based on parameter importance. For the residual vector The elements of each dimension are weighted and anomaly scores are calculated. ; S6. Based on the abnormal scoring results and residual vector weight allocation rules, diagnose the fault type and fault level, and verify the confidence level of the diagnosis results. S7 outputs diagnostic results and corresponding maintenance recommendations.

2. The diagnostic method for SCR catalyst blockage and deactivation according to claim 1, characterized in that, The multidimensional key parameters mentioned in S1 include the SCR system inlet and outlet NO. x Concentration, catalyst bed inlet and outlet pressure, ammonia slip concentration, SCR system operating temperature, and flue gas velocity.

3. The diagnostic method for SCR catalyst blockage and deactivation according to claim 1, characterized in that, The dynamic features for extracting multidimensional key parameters described in S3 include: S31, Extract the rate of change of pressure difference based on the trend of pressure difference change in the catalyst bed; S32, according to NO x The export concentration shows a long-term upward trend, and the extraction shows an increasing trend. S33, using the coefficient of variation to quantify the fluctuation characteristics of ammonia escape concentration, and extracting the coefficient of variation of ammonia escape; S34, extracting catalytic reaction efficiency indicators; S35, weighting and fusing the dynamic characteristics of multidimensional key parameters; S36, set threshold judgment conditions for the dynamic features of multidimensional key parameters, and perform feature threshold judgment.

4. The diagnostic method for SCR catalyst blockage and deactivation according to claim 3, characterized in that, In step S33, the mean and standard deviation of ammonia slip concentration are calculated hourly, and the coefficient of variation is calculated by the ratio of the mean to the standard deviation of ammonia slip concentration. This is represented by the ammonia slip concentration fluctuation characteristics, and the ammonia slip concentration fluctuation rate is calculated based on the ratio of the maximum and minimum differences in ammonia slip concentration. This can be represented using the following logic: In the formula, The coefficient of variation represents the ammonia escape concentration. This represents the standard deviation of ammonia slip concentration calculated from the time series. This represents the average ammonia slip concentration calculated from the time series. Indicates the ammonia escape concentration. Maximum ammonia escape concentration This represents the minimum ammonia escape concentration. This indicates the fluctuation rate of ammonia escape concentration.

5. The diagnostic method for SCR catalyst blockage and deactivation according to claim 3, characterized in that, S35 includes the following steps: S351, based on historical operating data and known system state labels, calculate the correlation score between each feature and the target state. ; S352, normalize the original relevance scores to obtain the final weight coefficients. ; S353, at each sampling time The feature vectors are weighted according to the weight coefficients to obtain the weighted feature vectors. This can be represented using the following logic: in, for The weighted feature vectors at each time step, For the first The eigenvectors at time... The value of , For the first Normalized weight coefficients of each eigenvector. Indicates the number of eigenvectors; S354 performs statistical and alignment processing on the weighted feature vectors according to a unified time window, specifically as follows: Set a fixed time window length The sampled data is segmented using the length of the time window as the sliding step. For the weighted feature data within each time window, the mean, variance, and moving average within that window are calculated sequentially to obtain the time window feature vector. : in, This represents the moving average operation. Indicates taking The mean, Indicates taking variance Indicates taking The moving average; The time window feature vectors of each feature are arranged sequentially in chronological order to form a time-series multidimensional weighted fusion feature matrix. : in, For multidimensional weighted fusion feature matrix, The number of time windows to look back. express The length of the time window.

6. The diagnostic method for SCR catalyst blockage and deactivation according to claim 5, characterized in that, The LSTM model described in S4 includes an input layer, an LSTM layer, a fully connected layer, and an output layer; The input layer is used to receive multidimensional feature matrices. ,in T represents the length of the time window. The dimension representing the multidimensional weighted fusion feature; The LSTM layer adopts a two-layer stacked structure, with 64 neurons in each layer and the activation function is tanh. The fully connected layer is used to map the final hidden state of the LSTM to a dimension of . The output space has the same dimensions as the input multidimensional feature vector; The output layer uses a linear activation function to generate the predicted value for the next time step. ; The training of the LSTM model specifically involves: using historical data from the SCR system under normal operating conditions as training data, employing Huber loss as the loss function, selecting the AdamW optimizer to optimize model parameters, validating model performance during training based on an early stopping strategy, suppressing overfitting through Dropout regularization, and representing the Huber loss using the following logic: In the formula, Indicates the loss value. Indicates the threshold parameter. express The multidimensional feature vectors observed at any given moment. express The predicted feature vector at time step.

7. The diagnostic method for SCR catalyst blockage and deactivation according to claim 1, characterized in that, S5 includes the following steps: S51, calculate based on the trained LSTM model Predicted value at time Compared with actual observation input The residual vector between This can be represented using the following logic: In the formula, for The multidimensional feature vectors observed at any given moment. , For the corresponding predicted feature vector, Let be the residual vector of the corresponding dimension, denoted as ; S52, assign weights based on parameter importance. For the residual vector The elements of each dimension are weighted and anomaly scores are calculated. This can be represented using the following logic: In the formula, For the first 3D residual vector ; S53, Adaptive dynamic threshold based on historical normal operating condition residual distribution To determine abnormal states, specifically: calculate the current residual score in real time. , and dynamic threshold When comparing, ≤ When the SCR system is in normal condition, it is determined that the system is in a normal state; when > When an anomaly is detected in the SCR system, step S6 is executed to further subdivide the anomaly level and assist in maintenance decision-making; wherein, the adaptive dynamic threshold... This represents the 95th percentile of the historical normal residual scores.

8. The diagnostic method for SCR catalyst blockage and deactivation according to claim 1, characterized in that, S6 includes the following steps: S61. Based on the contribution of different parameters in the residual vector, the fault type is distinguished using the following logical representation: S611, when the pressure difference change rate in S52 When the residual weight is >60% and the residual of the catalytic reaction efficiency η increases significantly, catalyst blockage is diagnosed. S612, when the ammonia escape variation coefficient in S52 When the residual weight is >50% and the catalytic reaction efficiency η continues to decrease, catalyst deactivation is diagnosed. S613, when S611 and S612 are satisfied at the same time, the catalyst is diagnosed to have both blockage and deactivation. S614, when neither S611 nor S612 is satisfied, the catalyst is diagnosed as normal; S62, quantify the fault level based on the anomaly scoring results, using the following logical representation: when And duration At that time, the catalyst state is determined to be either initial blockage or initial deactivation; when or and At that time, the catalyst condition is judged to be severely deactivated or severely blocked; Here, "Score" refers to the anomaly score calculated by the model based on the input feature vector at the current moment. This represents the initial threshold, calculated as the 90th percentile of the residual score from historical normal data. The severity threshold is represented by the 70th percentile of the residual score from historical fault data. Indicates the duration threshold. This indicates that the abnormal score has reached its first peak. The duration of the fault state; S63. Perform confidence verification on the diagnostic results and improve the reliability of the diagnostic results by introducing multi-indicator cross-validation.

9. The diagnostic method for SCR catalyst blockage and deactivation according to claim 1, characterized in that, Specifically, S7 is: The current catalyst status is output based on the model diagnostic results, including normal, initial blockage, severe blockage, and catalyst deactivation. When the output catalyst is in normal condition, no maintenance is recommended; when the output catalyst is in initial blockage condition, it is recommended to optimize the soot blowing frequency and check the soot blower; when the output catalyst is in severe blockage condition, it is recommended to shut down and flush with high pressure or switch to a backup module; when the output catalyst is in initial deactivation or severe deactivation condition, it is recommended to calibrate the ammonia injection control, evaluate regeneration, or replace it.

10. A diagnostic system for SCR catalyst blockage and deactivation, characterized in that, include: The data acquisition module is used to collect multi-dimensional key parameters of the SCR system in real time; The data preprocessing module is used to preprocess the collected multidimensional key parameters; The feature extraction module is used to extract dynamic features of multidimensional key parameters, perform weighted processing and alignment to form a multidimensional feature vector; wherein, the weighted processing and alignment to form a multidimensional feature vector includes: weighting and fusing the dynamic features of the multidimensional key parameters, specifically: Based on historical operational data and known system state labels, the correlation score between each feature and the target state is calculated. ; The original relevance scores are normalized to obtain the final weight coefficients. ; Weighted processing, specifically at each sampling time The feature vectors are weighted according to the weight coefficients to obtain the weighted feature vectors. ; Time window alignment and merging; The LSTM prediction module is used to build and train an LSTM model, taking a multi-dimensional feature vector as input and outputting the predicted value of the multi-dimensional feature vector at the next time step. The residual analysis module is used to calculate the residual vectors between the input and output, and to generate anomaly scores based on weighted Euclidean distance. Specifically: Calculate the residual vector ; Weighted Euclidean distance is used to enhance the sensitivity of key features to anomalies, specifically by assigning weights based on parameter importance. For the residual vector The elements of each dimension are weighted and anomaly scores are calculated. ; The status classification module is used to diagnose fault types and fault levels based on anomaly scoring results and residual vector weight allocation rules, and to verify the confidence level of the diagnostic results. The results output module is used to output diagnostic results and corresponding maintenance suggestions.

Images