Flue gas desulfurization equipment health state monitoring method and system based on deep neural network

Abstract

The invention discloses a flue gas desulfurization equipment health state monitoring method and system based on a deep neural network, and belongs to the field of flue gas desulfurization equipment health state monitoring, and the method comprises the steps: obtaining the operation data of flue gas desulfurization equipment, and carrying out the denoising processing based on the noise data through wavelet transform, and obtaining the denoised operation data; extracting multi-dimensional characteristics of inlet flue gas temperature, sulfur dioxide concentration and absorbent flow from the operation data after precise cleaning, and constructing a characteristic matrix; determining an optimal network structure of the deep neural network by adopting Bayesian optimization according to the feature matrix to obtain an optimized network structure; and constructing a deep neural network model by adopting the optimized network structure, training the model through a back propagation algorithm, and if a training error is lower than a preset threshold, stopping training to obtain a trained model.

Description

Technical Field

[0001] The present invention belongs to the field of flue gas desulfurization equipment health status monitoring, and in particular, relates to a flue gas desulfurization equipment health status monitoring method and system based on a deep neural network. Background Art

[0002] In the actual operation environment of flue gas desulfurization equipment, the collected data often contains a lot of noise and outliers. Directly using it for the training of deep neural network models will lead to a decline in model performance. The data cleaning process needs to remove abnormal data, but under certain operating conditions, such as startup and shutdown or drastic load changes, the desulfurization equipment will fluctuate within the normal range. These fluctuating data are highly similar to abnormal data in certain features. How to accurately distinguish normal fluctuations from real equipment abnormalities is a difficult problem. Furthermore, when building a deep neural network model, the number of hidden layers, the number of neurons in each layer, and the choice of activation function will directly affect the performance of the model.

[0003] In the scenario of health status monitoring of flue gas desulfurization equipment, the operating status of the equipment is affected by many factors, including inlet flue gas temperature, sulfur dioxide concentration, absorbent flow rate, etc., and there are complex nonlinear relationships between these factors. How to determine an optimal network structure to accurately capture these complex nonlinear relationships and avoid overfitting or underfitting of the model is a key issue. In addition, the operating status of the flue gas desulfurization equipment is a dynamic process that changes over time, and historical data has an important impact on the current status. Although the introduction of the long short-term memory layer can capture the long-term dependencies in time series data, it has many internal parameters and a complex training process. How to determine the appropriate number of memory units and the forget gate threshold to balance the long-term memory capacity of the model and real-time computing efficiency is another technical problem that needs to be solved. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a method for monitoring the health status of flue gas desulfurization equipment based on a deep neural network, comprising: Acquire flue gas desulfurization equipment operation data, and perform wavelet transform denoising on noise data of the flue gas desulfurization equipment operation data to obtain denoised operation data; The denoised operation data is tested by an isolation forest algorithm, and if a data point in the operation data is determined to be abnormal, the abnormal data point is removed to obtain the operation data after preliminary cleaning; A dynamic time warping algorithm is used to calculate the characteristic similarity between normal operating condition fluctuations and equipment abnormalities in the operation data after the preliminary cleaning, so as to obtain the operation data after accurate cleaning; Extracting multi-dimensional features of inlet flue gas temperature, sulfur dioxide concentration, and absorbent flow rate based on the operation data after precise cleaning, and constructing a feature matrix; According to the feature matrix, using Bayesian optimization to determine the optimal network structure of the deep neural network to obtain an optimized network structure; A deep neural network model is constructed using the optimized network structure, and the model is trained using a back propagation algorithm. If the training error is lower than a preset threshold, the training is stopped to obtain a trained model. Introducing a long short-term memory layer into the trained model, determining the number of memory units according to the length of historical data, and increasing the number of memory units if the length of historical data is greater than a set value, otherwise reducing the number of memory units to obtain a dynamically adjusted model; Based on the dynamically adjusted model, the operating status of the flue gas desulfurization equipment is monitored in real time. If the deviation between the monitoring result and the preset status exceeds a threshold, an early warning mechanism is triggered to obtain the equipment health status monitoring result.

[0005] Preferably, the process of obtaining the denoised operating data includes: Acquire flue gas desulfurization equipment operation data, and perform wavelet transform denoising on noise data of the flue gas desulfurization equipment operation data to obtain denoised data; According to the characteristics of the denoised data, determine whether the data quality meets the preset standards. If the data quality meets the standards, use the support vector machine algorithm to classify the data. If the data quality does not meet the standards, use the principal component analysis algorithm to reduce the dimension of the data. According to the data characteristics after dimensionality reduction, the random forest algorithm is used to perform regression analysis on the data. The regression analysis results are used to determine whether the data accuracy reaches the preset threshold. If the data accuracy reaches the threshold, the final processing result is output.

[0006] Preferably, the process of obtaining the operation data after preliminary cleaning includes: Constructing an isolation forest model, training the isolation forest model, and obtaining a trained isolation forest model; Based on the trained isolation forest model, the denoised operating data is tested, and if a data point in the operating data is determined to be an abnormal data point, the data point is marked to obtain marked operating data; According to the marks of the data points in the operation data, the marked data points are removed to obtain the operation data after the abnormal data points are removed, and according to the operation data after the abnormal data points are removed, the number of the data points is counted, and if the number is lower than a preset threshold, the data points are supplemented to obtain the supplemented operation data; According to the supplemented operation data, the data point density is calculated. If the data point density is lower than a preset threshold, data point sampling is performed to obtain the sampled operation data. According to the sampled operating data, the variance of the data point is calculated. If the variance is greater than a preset threshold, the data point is smoothed to obtain the smoothed operating data. According to the smoothed operating data, a data feature vector is constructed, and a principal component analysis algorithm is used to perform dimensionality reduction processing to obtain the operating data after preliminary cleaning.

[0007] Preferably, the process of obtaining the operation data after precise cleaning includes: Constructing a time series matrix based on the running data of the initial state; wherein each row of the time series matrix represents all data point values ​​at a time point; The sliding window method is used to segment the time series matrix to obtain multiple time series sub-windows; Based on each time series sub-window, statistical features are extracted to obtain the statistical feature vector of each sub-window; According to the statistical eigenvector, the normal operating condition fluctuation range of each time series sub-window is calculated; Obtain the abnormal operation data of the equipment, extract its statistical features, and obtain the feature vector of the equipment abnormality; According to the feature vector of equipment anomaly and the fluctuation range of normal working condition, the dynamic time warping algorithm is used to calculate the feature similarity between each time series sub-window and the equipment anomaly, and the feature similarity value of each time series sub-window is obtained; According to the feature similarity value of each time series sub-window, it is judged whether it is greater than the set threshold. If it is greater than the set threshold, the data points corresponding to the time series sub-window are retained, otherwise the data points corresponding to the time series sub-window are removed to obtain the accurately cleaned running data.

[0008] Preferably, the process of constructing the feature matrix includes: According to the operation data after precise cleaning, the multi-dimensional characteristic data of inlet flue gas temperature, sulfur dioxide concentration and absorbent flow rate are obtained; Based on the acquired multi-dimensional feature data, a data standardization method is used to normalize the feature data to obtain normalized multi-dimensional feature data; The normalized multi-dimensional feature data is reduced in dimension by principal component analysis algorithm, and the principal component that best represents the original data characteristics is extracted as a new feature; According to the results of principal component analysis, the first k principal components whose cumulative variance contribution rate reaches the preset threshold are selected to construct the feature matrix after dimensionality reduction.

[0009] Preferably, according to the feature matrix, the optimal network structure of the deep neural network is determined by using Bayesian optimization, and the process of obtaining the optimized network structure includes: Acquire different sample categories based on the feature matrix; Build a classification model based on the decision tree algorithm and determine the optimal decision tree parameters through cross-validation; The classification results of the decision tree classification model are used as prior probability and combined with the Bayesian optimization algorithm to search for the optimal network structure parameters of the deep neural network.

[0010] Preferably, the process of obtaining the trained model includes: Obtain the training data set, use the optimized network structure to build a multi-level deep neural network model, and set appropriate hyperparameters and activation functions; The training data set is input into the constructed deep neural network model, and the back propagation algorithm is used to adaptively adjust the weights and thresholds of neurons in each layer to continuously optimize the model parameters; During each iterative training process, the training error of the current model is calculated. If the training error is lower than the preset threshold, the model training is determined to be complete, the training process is stopped, and a trained model is obtained.

[0011] Preferably, the process of obtaining the dynamically adjusted model includes: Get the trained model, introduce the long short-term memory layer, and build a dynamic optimization model; Obtain the historical data to be processed and determine the historical length of the data; Determine whether the length of historical data exceeds a preset threshold. If so, dynamically increase the number of memory units. If the threshold is not exceeded, the number of memory cells is dynamically reduced; Reconfigure the structure and parameters of the long short-term memory layer according to the adjusted number of memory units; The reconfigured long short-term memory layer is embedded into the initial model to obtain an optimized model that introduces a dynamic memory mechanism.

[0012] On the other hand, the present invention also provides a flue gas desulfurization equipment health status monitoring system based on a deep neural network, comprising: Data acquisition module, used to obtain flue gas desulfurization equipment operation data; A denoising processing module, used for performing denoising processing based on the noise data of the flue gas desulfurization equipment operation data by using wavelet transform to obtain denoised operation data; An anomaly detection module is used to detect the denoised operation data by using an isolation forest algorithm. If a data point in the operation data is determined to be abnormal, the abnormal data point is removed to obtain the operation data after preliminary cleaning; A data cleaning module, used to calculate the characteristic similarity between normal operating condition fluctuations and equipment abnormalities in the operation data after the preliminary cleaning by using a dynamic time warping algorithm, so as to obtain the operation data after accurate cleaning; A feature extraction module, used to extract multi-dimensional features of inlet flue gas temperature, sulfur dioxide concentration, and absorbent flow rate based on the operation data after the precise cleaning, and construct a feature matrix; A network optimization module, used to determine the optimal network structure of the deep neural network using Bayesian optimization according to the feature matrix to obtain an optimized network structure; A model training module is used to construct a deep neural network model using the optimized network structure, train the model using a back propagation algorithm, and stop training if the training error is lower than a preset threshold to obtain a trained model; The status monitoring module is used to introduce a long and short-term memory layer into the trained model, determine the number of memory units according to the length of historical data, increase the number of memory units if the length of historical data is greater than a set value, otherwise reduce the number of memory units to obtain a dynamically adjusted model; based on the dynamically adjusted model, the operating status of the flue gas desulfurization equipment is monitored in real time. If the deviation between the monitoring result and the preset status exceeds a threshold, an early warning mechanism is triggered to obtain the equipment health status monitoring result.

[0013] Compared with the prior art, the present invention has the following advantages and technical effects: The present invention discloses a method for monitoring the health status of flue gas desulfurization equipment based on a deep neural network. The method first performs wavelet transform denoising and anomaly detection on the equipment operation data, and then uses a dynamic time warping algorithm to perform accurate data cleaning. Then, multidimensional features are extracted from the cleaned data to construct a feature matrix. Based on this feature matrix, Bayesian optimization is used to determine the optimal structure of the deep neural network, and a long and short-term memory layer is introduced to dynamically adjust the number of memory units according to the length of historical data. Finally, the dynamically adjusted model is used to monitor the operating status of the flue gas desulfurization equipment in real time, and an early warning mechanism is triggered when the deviation between the monitoring result and the preset state exceeds a threshold. The present invention realizes accurate monitoring and timely early warning of the operating status of the flue gas desulfurization equipment through the combination of multiple data processing and intelligent algorithms, effectively improving the safety and stability of the equipment operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings: Figure 1 It is a flow chart of a method for monitoring the health status of flue gas desulfurization equipment based on a deep neural network according to an embodiment of the present invention; Figure 2 Schematic diagram of the system structure of an embodiment of the present invention. DETAILED DESCRIPTION

[0015] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0016] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0017] Embodiment 1

[0018] like Figure 1 As shown, this embodiment provides a method for monitoring the health status of flue gas desulfurization equipment based on a deep neural network, including: S101, obtaining flue gas desulfurization equipment operation data, and performing wavelet transform denoising on noise data to obtain denoised operation data.

[0019] Obtain the operating data of the flue gas desulfurization equipment, use wavelet transform to denoise the noise part in the data, and obtain the denoised operating data. According to the characteristics of the denoised data, determine whether the data quality meets the preset standards. If the data quality meets the standards, use the support vector machine algorithm to classify the data. If the data quality does not meet the standards, use the principal component analysis algorithm to reduce the dimension of the data. According to the characteristics of the data after dimensionality reduction, use the random forest algorithm to perform regression analysis on the data. Through the regression analysis results, determine whether the data accuracy reaches the preset threshold. If the data accuracy reaches the threshold, output the final processing result. If the data accuracy does not reach the threshold, re-acquire the data and perform a new round of processing.

[0020] Specifically, the acquisition of flue gas desulfurization equipment operation data is an important part of environmental monitoring. These data usually include key indicators such as sulfur dioxide concentration, pH value, liquid-gas ratio, etc. However, the original data often has noise interference, which affects the data quality and analysis accuracy. Wavelet transform is an effective denoising method, which can analyze signals in the time and frequency domains simultaneously, retain useful information and remove noise. For example, for sulfur dioxide concentration data, the db4 wavelet basis function can be selected, 5-layer decomposition, and soft threshold method denoising can be used to effectively eliminate high-frequency noise and retain low-frequency trends. The quality assessment of the denoised data is a key step to ensure the reliability of subsequent analysis. The preset standards may include aspects such as data integrity, consistency, and accuracy. If the data quality meets the standards, the support vector machine (SVM) algorithm can be used for data classification. For example, the desulfurization efficiency can be classified according to different operating conditions (such as load level, fuel type), a multi-class SVM model can be established, the kernel function parameters can be optimized, and the classification accuracy can be improved. When the data quality does not meet the standards, principal component analysis (PCA) can be used for dimensionality reduction. PCA can extract the main features in the data and reduce redundant information. In the flue gas desulfurization system, there may be multiple parameters with strong correlation, such as absorption tower pressure, slurry density, etc. Through PCA, these high-dimensional data can be reduced to 2 to 3 principal components, retaining more than 80% of the information, simplifying subsequent analysis. The random forest algorithm is a powerful ensemble learning method suitable for regression analysis. In the prediction of desulfurization efficiency, influencing factors such as inlet sulfur dioxide concentration, absorption tower pH value, liquid-gas ratio, etc. can be selected as feature variables to establish a random forest model. By setting the appropriate number of trees (such as 500 trees) and the minimum number of node samples, stable and reliable prediction results can be obtained. Data accuracy assessment is an important part of model application. The prediction error can be set to less than 5% as the threshold standard. If the threshold is reached, it means that the model performance is good, and the final processing result can be output to guide the optimized operation of the desulfurization equipment. If the threshold is not reached, the data needs to be acquired again, and the sampling frequency, sensor position, etc. may need to be adjusted to improve the data quality.

[0021] S102, using an isolation forest algorithm to detect abnormal data in the denoised operating data, if a data point in the operating data is determined to be abnormal, removing the abnormal data point to obtain preliminarily cleaned operating data.

[0022] Obtain denoised operating data, train a model based on the isolation forest algorithm, and obtain a trained isolation forest model. Apply the trained isolation forest model to detect the operating data. If a data point in the operating data is determined to be an abnormal data point, mark the data point to obtain the marked operating data. According to the marking of the data point in the operating data, remove the marked data point to obtain the operating data after removing the abnormal data point. According to the operating data after removing the abnormal data point, calculate the number of data points. If the number is lower than a preset threshold, supplement the data points to obtain the supplemented operating data. According to the supplemented operating data, calculate the data point density. If the data point density is lower than a preset threshold, sample the data points to obtain the sampled operating data. According to the sampled operating data, calculate the data point variance. If the variance is greater than a preset threshold, smooth the data points to obtain the smoothed operating data. According to the smoothed operating data, construct a data feature vector, use the principal component analysis algorithm to perform dimensionality reduction, and obtain the preliminary cleaned operating data.

[0023] Specifically, the isolation forest algorithm is an anomaly detection method that identifies anomalies in a data set by building a decision tree. In the processing of flue gas desulfurization equipment operation data, the application of this algorithm can effectively detect abnormal data. For example, the pH value of a desulfurization equipment is usually between 6.5 and 7.5. If a data point with a pH value of 3 or 10 appears, the isolation forest model will mark it as an anomaly. After marking the abnormal data points, these points need to be removed to ensure data quality. Assume that among 1000 data points, 20 are marked as abnormal, and 980 valid data points remain after removal. If the preset threshold is 950, there is no need to add data points; but if the threshold is 1000, 20 data points need to be added to meet the requirements. The density of data points reflects the uniformity of data distribution. Taking the desulfurization efficiency data as an example, if the data points are mainly concentrated in the 40% to 60% interval within the range of 0% to 100%, and the data in other intervals are sparse, sampling may be required to increase the overall data density. The sampling method can be simple random sampling or stratified sampling to ensure that the number of data points in each interval is balanced. The variance of the data points reflects the degree of discreteness of the data. For flue gas flow data, if the variance is too large, it may indicate that the equipment is unstable or there is a measurement error. At this time, smoothing processing is required, such as using the moving average method to reduce abnormal fluctuations and make the data smoother. Feature vector construction is a key step in data analysis. For flue gas desulfurization systems, inlet SO2 concentration, desulfurizer dosage, desulfurization efficiency, etc. can be selected as features. Principal component analysis (PCA) can reduce the data dimension and extract the most important features. For example, reducing 10 features to 3-4 principal components not only retains most of the information, but also simplifies the subsequent analysis process. Through this series of processing, higher quality and more representative operating data are obtained.

[0024] S103. For the operation data after the preliminary cleaning, a dynamic time warping algorithm is used to calculate the feature similarity between normal operating condition fluctuations and equipment abnormalities. If the feature similarity is higher than a preset threshold, the data point is retained, otherwise the data point is discarded to obtain the operation data after accurate cleaning.

[0025] According to the operation data of the initial state, a time series matrix is ​​constructed. Each row of the time series matrix represents all the data point values ​​at a time point. The sliding window method is used to segment the time series matrix to obtain multiple time series sub-windows, each of which contains the operation data in a continuous time. For each time series sub-window, statistical features are extracted to obtain the statistical feature vector of each sub-window. The statistical feature vector contains the mean and variance. According to the statistical feature vector, the normal operating condition fluctuation range of each time series sub-window is calculated. The normal operating condition fluctuation range is a numerical interval. The operation data of the equipment abnormality is obtained, and its statistical features are extracted to obtain the feature vector of the equipment abnormality. The feature vector of the equipment abnormality contains the mean and variance. According to the feature vector of the equipment abnormality and the normal operating condition fluctuation range, the dynamic time warping algorithm is used to calculate the feature similarity of each time series sub-window with the equipment abnormality, and the feature similarity value of each time series sub-window is obtained. According to the feature similarity value of each time series sub-window, it is judged whether it is greater than the set threshold. If it is greater than the set threshold, the data points corresponding to the time series sub-window are retained, otherwise the data points corresponding to the time series sub-window are eliminated, and finally the operation data after accurate cleaning is obtained.

[0026] S104, extracting multi-dimensional features of inlet flue gas temperature, sulfur dioxide concentration, and absorbent flow rate from the operation data after precise cleaning, and constructing a feature matrix.

[0027] According to the operation data after precise cleaning, multi-dimensional feature data such as inlet flue gas temperature, sulfur dioxide concentration and absorbent flow rate are obtained. For the obtained multi-dimensional feature data, the data standardization method is used to normalize the feature data to eliminate the dimensional differences between different features. The normalized multi-dimensional feature data is reduced in dimension by the principal component analysis algorithm, and the principal component that best represents the original data characteristics is extracted as a new feature. According to the results of the principal component analysis, the first k principal components whose cumulative variance contribution rate reaches the preset threshold are selected to construct the feature matrix after dimension reduction. The K-means clustering algorithm is used to perform cluster analysis on the feature matrix after dimension reduction, and the data is divided into different operating conditions according to the clustering results. For each operating condition, the decision tree algorithm is used to establish a nonlinear relationship model between the inlet flue gas temperature, sulfur dioxide concentration and absorbent flow rate and the desulfurization efficiency. The established desulfurization efficiency prediction model is applied to the actual production process, and the desulfurization efficiency is predicted according to the inlet flue gas parameters collected in real time, providing a basis for optimization control.

[0028] Specifically, in the desulfurization system, the operation data after accurate cleaning lays the foundation for subsequent analysis. First, multidimensional feature data such as inlet flue gas temperature, sulfur dioxide concentration and absorbent flow rate are obtained. These feature data reflect the key operating parameters of the desulfurization system and directly affect the desulfurization efficiency. For example, the inlet flue gas temperature may fluctuate between 120℃ and 180℃, the sulfur dioxide concentration may vary between 2000mg / m³ and 5000mg / m³, and the absorbent flow rate may be adjusted between 50m³ / h and 100m³ / h. In order to eliminate the dimensional differences between different features, the data standardization method is used for normalization. This step ensures that features of different dimensions can be compared and analyzed on the same scale. For example, through z-score standardization, each feature can be converted into a standard normal distribution with a mean of 0 and a standard deviation of 1. Next, the principal component analysis (PCA) algorithm is used to reduce the dimension of the normalized multidimensional feature data. PCA can extract the principal components that best represent the characteristics of the original data, effectively reducing the data dimension while retaining key information. Assuming that the first three principal components with a cumulative variance contribution rate of 95% are selected, these principal components may represent potential factors such as the combined effect of temperature-concentration, absorbent utilization efficiency, and system stability. After dimensionality reduction, the K-means clustering algorithm is used to perform cluster analysis on the feature matrix. This step can divide the data into different operating conditions, such as high load, medium load, and low load conditions. Under each operating condition, the operating characteristics and parameter relationships of the desulfurization system may be significantly different. For each operating condition, a decision tree algorithm is used to establish a nonlinear relationship model between the inlet flue gas parameters and the desulfurization efficiency.

[0029] S105. According to the feature matrix, Bayesian optimization is used to determine the optimal network structure of the deep neural network to obtain an optimized network structure.

[0030] According to the feature matrix, the principal component analysis method is used to reduce the dimension of the features to obtain the feature subset after dimensionality reduction; for the feature subset, the K-means clustering algorithm is used to cluster the data to obtain different sample categories; according to the clustered sample categories, the decision tree algorithm is used to build a classification model, and the optimal decision tree parameters are determined through cross-validation; the classification results of the decision tree classification model are used as prior probabilities, and the optimal network structure parameters of the deep neural network are searched in combination with the Bayesian optimization algorithm; in the Bayesian optimization search process, the neural network performance under the current network structure parameters is used as the optimization objective function, and the optimal network structure is obtained by maximizing the objective function; according to the optimal network structure parameters obtained by Bayesian optimization, the corresponding deep neural network model is constructed, and the training is performed on the training data to obtain the final network model; the performance of the trained deep neural network model is evaluated using the test data, and the accuracy, recall rate and other indicators of the model are calculated to verify the generalization ability of the model.

[0031] S106. Construct a deep neural network model using the optimized network structure, train the model using a back propagation algorithm, and stop training if the training error is lower than a preset threshold to obtain a trained model.

[0032] According to the pre-established business knowledge base, the data and rules related to the current business scenario are obtained, and feature extraction and data preprocessing are performed to obtain a standardized training data set. For the training data set, the optimized network structure is used to build a multi-level deep neural network model, and appropriate hyperparameters and activation functions are set. The training data set is input into the constructed deep neural network model, and the back propagation algorithm is used to adaptively adjust the weights and thresholds of each layer of neurons to continuously optimize the model parameters. In each iterative training process, the training error of the current model is calculated. If the training error is lower than the preset threshold, the model training is determined to be completed and the training process is stopped. The trained deep neural network model is obtained, and the corresponding output results are obtained for the newly input business data through the forward propagation calculation of the model. The output results of the deep neural network model are compared with the expected target, and its deviation is calculated. According to the deviation, the model is fine-tuned and the parameters are optimized.

[0033] S107. Introduce a long short-term memory layer into the trained model, determine the number of memory units according to the length of historical data, and if the length of historical data is greater than a set value, increase the number of memory units; otherwise, reduce the number of memory units to obtain a dynamically adjusted model.

[0034] Obtain the trained initial model, introduce the long and short-term memory layer, and build a dynamic optimization model. Obtain the historical data to be processed and determine the historical length of the data. Determine whether the length of the historical data exceeds the preset threshold. If it exceeds the threshold, dynamically increase the number of memory units; if it does not exceed the threshold, dynamically reduce the number of memory units. Reconfigure the structure and parameters of the long and short-term memory layer according to the adjusted number of memory units. Embed the reconfigured long and short-term memory layer into the initial model to obtain an optimized model that introduces a dynamic memory mechanism. Use the optimized model to process and predict new input data, dynamically adjust the model's memory capacity according to data characteristics, and improve the model's adaptability and accuracy. Continuously monitor the changes in the historical length of the input data, dynamically adjust the number of memory units according to the length change trend, realize the adaptive optimization of the model, and maintain the stability and improvement of the model's performance.

[0035] S108. Use the dynamically adjusted model to monitor the operating status of the flue gas desulfurization equipment in real time. If the deviation between the monitoring result and the preset status exceeds a threshold, an early warning mechanism is triggered to obtain the equipment health status monitoring result.

[0036] Based on the historical operation data of the flue gas desulfurization equipment, a machine learning algorithm is used to establish an equipment operation status prediction model to obtain dynamically adjusted model parameters; the dynamically adjusted model is used to analyze the real-time operation data of the flue gas desulfurization equipment to obtain the monitoring results of the current operation status of the equipment; the monitoring results are compared with the preset normal operation status of the equipment, the deviation value is calculated, and it is determined whether the deviation value exceeds the preset threshold; if the deviation value exceeds the preset threshold, the early warning mechanism is triggered and an equipment abnormality early warning signal is sent to the system; based on the early warning signal, the equipment fault diagnosis knowledge base is obtained, and the fault diagnosis algorithm is used to analyze the current operation data of the equipment to determine the potential fault type and cause of the equipment; according to the diagnosed equipment fault type and cause, the corresponding treatment measures are matched from the preset fault solution knowledge base to generate an equipment maintenance plan; the equipment health status monitoring results, fault diagnosis results and maintenance plans are output and displayed to provide equipment management decision-making basis for operation and maintenance personnel.

[0037] like Figure 2 As shown, this embodiment also provides a flue gas desulfurization equipment health status monitoring system based on a deep neural network, including: Data acquisition module, used to obtain flue gas desulfurization equipment operation data; A denoising processing module, used for performing denoising processing based on the noise data of the flue gas desulfurization equipment operation data by using wavelet transform to obtain denoised operation data; An anomaly detection module is used to detect the denoised operation data by using an isolation forest algorithm. If a data point in the operation data is determined to be abnormal, the abnormal data point is removed to obtain the operation data after preliminary cleaning; A data cleaning module, used to calculate the characteristic similarity between normal operating condition fluctuations and equipment abnormalities in the operation data after the preliminary cleaning by using a dynamic time warping algorithm, so as to obtain the operation data after accurate cleaning; A feature extraction module, used to extract multi-dimensional features of inlet flue gas temperature, sulfur dioxide concentration, and absorbent flow rate based on the operation data after the precise cleaning, and construct a feature matrix; A network optimization module, used to determine the optimal network structure of the deep neural network using Bayesian optimization according to the feature matrix to obtain an optimized network structure; A model training module is used to construct a deep neural network model using the optimized network structure, train the model using a back propagation algorithm, and stop training if the training error is lower than a preset threshold to obtain a trained model; The status monitoring module is used to introduce a long and short-term memory layer into the trained model, determine the number of memory units according to the length of historical data, increase the number of memory units if the length of historical data is greater than a set value, otherwise reduce the number of memory units to obtain a dynamically adjusted model; based on the dynamically adjusted model, the operating status of the flue gas desulfurization equipment is monitored in real time. If the deviation between the monitoring result and the preset status exceeds a threshold, an early warning mechanism is triggered to obtain the equipment health status monitoring result.

[0038] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for monitoring the health status of flue gas desulfurization equipment based on deep neural network, characterized in that: include: Acquire flue gas desulfurization equipment operation data, and perform wavelet transform denoising on noise data of the flue gas desulfurization equipment operation data to obtain denoised operation data; The denoised operation data is tested by an isolation forest algorithm, and if a data point in the operation data is determined to be abnormal, the abnormal data point is removed to obtain the operation data after preliminary cleaning; A dynamic time warping algorithm is used to calculate the characteristic similarity between normal operating condition fluctuations and equipment abnormalities in the operation data after the preliminary cleaning, so as to obtain accurate operation data after cleaning; Extracting multi-dimensional features of inlet flue gas temperature, sulfur dioxide concentration, and absorbent flow rate based on the operation data after precise cleaning, and constructing a feature matrix; According to the feature matrix, using Bayesian optimization to determine the optimal network structure of the deep neural network to obtain an optimized network structure; A deep neural network model is constructed using the optimized network structure, and the model is trained using a back propagation algorithm. If the training error is lower than a preset threshold, the training is stopped to obtain a trained model. Introducing a long short-term memory layer into the trained model, determining the number of memory units according to the length of historical data, and increasing the number of memory units if the length of historical data is greater than a set value, otherwise reducing the number of memory units to obtain a dynamically adjusted model; Based on the dynamically adjusted model, the operating status of the flue gas desulfurization equipment is monitored in real time. If the deviation between the monitoring result and the preset status exceeds a threshold, an early warning mechanism is triggered to obtain the equipment health status monitoring result.

2. The method according to claim 1, characterized in that The process of obtaining the denoised operating data includes: Acquire flue gas desulfurization equipment operation data, and perform wavelet transform denoising on noise data of the flue gas desulfurization equipment operation data to obtain denoised data; According to the characteristics of the denoised data, determine whether the data quality meets the preset standards. If the data quality meets the standards, use the support vector machine algorithm to classify the data. If the data quality does not meet the standards, use the principal component analysis algorithm to reduce the dimension of the data. According to the data characteristics after dimensionality reduction, the random forest algorithm is used to perform regression analysis on the data. The regression analysis results are used to determine whether the data accuracy reaches the preset threshold. If the data accuracy reaches the threshold, the final processing result is output.

3. The method according to claim 1, characterized in that The process of obtaining the operation data after preliminary cleaning includes: Constructing an isolation forest model, training the isolation forest model, and obtaining a trained isolation forest model; Based on the trained isolation forest model, the denoised operating data is tested, and if a data point in the operating data is determined to be an abnormal data point, the data point is marked to obtain marked operating data; According to the marks of the data points in the operation data, the marked data points are removed to obtain the operation data after the abnormal data points are removed, and according to the operation data after the abnormal data points are removed, the number of the data points is counted, and if the number is lower than a preset threshold, the data points are supplemented to obtain the supplemented operation data; According to the supplemented operation data, the data point density is calculated. If the data point density is lower than a preset threshold, data point sampling is performed to obtain the sampled operation data. According to the sampled operating data, the variance of the data point is calculated. If the variance is greater than a preset threshold, the data point is smoothed to obtain the smoothed operating data. According to the smoothed operating data, a data feature vector is constructed, and a principal component analysis algorithm is used to perform dimensionality reduction processing to obtain the operating data after preliminary cleaning.

4. The method according to claim 1, characterized in that: The process of obtaining the operation data after precise cleaning includes: Constructing a time series matrix based on the running data of the initial state; wherein each row of the time series matrix represents all data point values ​​at a time point; The sliding window method is used to segment the time series matrix to obtain multiple time series sub-windows; Based on each time series sub-window, statistical features are extracted to obtain the statistical feature vector of each sub-window; According to the statistical eigenvector, the normal operating condition fluctuation range of each time series sub-window is calculated; Obtain the abnormal operation data of the equipment, extract its statistical features, and obtain the feature vector of the equipment abnormality; According to the feature vector of equipment anomaly and the fluctuation range of normal working condition, the dynamic time warping algorithm is used to calculate the feature similarity between each time series sub-window and the equipment anomaly, and the feature similarity value of each time series sub-window is obtained; According to the feature similarity value of each time series sub-window, it is judged whether it is greater than the set threshold. If it is greater than the set threshold, the data points corresponding to the time series sub-window are retained, otherwise the data points corresponding to the time series sub-window are removed to obtain the accurately cleaned running data.

5. The method according to claim 1, characterized in that The process of constructing the feature matrix includes: According to the operation data after precise cleaning, the multi-dimensional characteristic data of inlet flue gas temperature, sulfur dioxide concentration and absorbent flow rate are obtained; Based on the acquired multi-dimensional feature data, a data standardization method is used to normalize the feature data to obtain normalized multi-dimensional feature data; The normalized multi-dimensional feature data is reduced in dimension by principal component analysis algorithm, and the principal component that best represents the original data characteristics is extracted as a new feature; According to the results of principal component analysis, the first k principal components whose cumulative variance contribution rate reaches the preset threshold are selected to construct the feature matrix after dimensionality reduction.

6. The method according to claim 1, characterized in that According to the feature matrix, the optimal network structure of the deep neural network is determined by using Bayesian optimization. The process of obtaining the optimized network structure includes: Acquire different sample categories based on the feature matrix; Build a classification model based on the decision tree algorithm and determine the optimal decision tree parameters through cross-validation; The classification results of the decision tree classification model are used as prior probability and combined with the Bayesian optimization algorithm to search for the optimal network structure parameters of the deep neural network.

7. The method according to claim 1, characterized in that The process of obtaining the trained model includes: Obtain the training data set, use the optimized network structure to build a multi-level deep neural network model, and set appropriate hyperparameters and activation functions; The training data set is input into the constructed deep neural network model, and the back propagation algorithm is used to adaptively adjust the weights and thresholds of neurons in each layer to continuously optimize the model parameters; During each iterative training process, the training error of the current model is calculated. If the training error is lower than the preset threshold, the model training is determined to be complete, the training process is stopped, and a trained model is obtained.

8. The method according to claim 1, characterized in that The process of obtaining the dynamically adjusted model includes: Get the trained model, introduce the long short-term memory layer, and build a dynamic optimization model; Obtain the historical data to be processed and determine the historical length of the data; Determine whether the length of historical data exceeds a preset threshold. If so, dynamically increase the number of memory units. If the threshold is not exceeded, the number of memory cells is dynamically reduced; Reconfigure the structure and parameters of the long short-term memory layer according to the adjusted number of memory units; The reconfigured long short-term memory layer is embedded into the initial model to obtain an optimized model that introduces a dynamic memory mechanism.

9. A flue gas desulfurization equipment health status monitoring system based on deep neural network, characterized in that: include: Data acquisition module, used to obtain flue gas desulfurization equipment operation data; A denoising processing module, used for performing denoising processing based on the noise data of the flue gas desulfurization equipment operation data by using wavelet transform to obtain denoised operation data; An anomaly detection module is used to detect the denoised operation data by using an isolation forest algorithm. If a data point in the operation data is determined to be abnormal, the abnormal data point is removed to obtain the operation data after preliminary cleaning; A data cleaning module, used to calculate the characteristic similarity between normal operating condition fluctuations and equipment abnormalities in the operation data after the preliminary cleaning by using a dynamic time warping algorithm, so as to obtain the operation data after accurate cleaning; A feature extraction module, used to extract multi-dimensional features of inlet flue gas temperature, sulfur dioxide concentration, and absorbent flow rate based on the operation data after the precise cleaning, and construct a feature matrix; A network optimization module, used to determine the optimal network structure of the deep neural network using Bayesian optimization according to the feature matrix to obtain an optimized network structure; A model training module is used to construct a deep neural network model using the optimized network structure, train the model using a back propagation algorithm, and stop training if the training error is lower than a preset threshold to obtain a trained model; The status monitoring module is used to introduce a long and short-term memory layer into the trained model, determine the number of memory units according to the length of historical data, increase the number of memory units if the length of historical data is greater than a set value, otherwise reduce the number of memory units to obtain a dynamically adjusted model; based on the dynamically adjusted model, the operating status of the flue gas desulfurization equipment is monitored in real time. If the deviation between the monitoring result and the preset status exceeds a threshold, an early warning mechanism is triggered to obtain the equipment health status monitoring result.

Images