Flue gas desulfurization process-oriented data cross-modal characterization and modeling method
By combining hierarchical models, temporal correlation models, and deep autoencoders, the problem of chaotic cross-modal data attribution during flue gas desulfurization was solved. Low-dimensional feature vectors were extracted, and a fuzzy neural network model was constructed to achieve efficient characterization of the desulfurization process.
Patent Information
- Application Number
- CN202511392844.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-27
- Publication Date
- 2025-11-07
AI Technical Summary
In existing technologies, cross-modal data processing in flue gas desulfurization processes leads to chaotic data attribution relationships, making it impossible to effectively mine the interaction patterns of cross-equipment data and the continuity characteristics of cross-stage data, and making it difficult to reflect the overall operating status of the desulfurization process.
Hierarchical models and temporal correlation models are used to clarify the hierarchical relationship of cross-modal data. Low-dimensional feature vectors are extracted by deep autoencoders, and a fuzzy neural network model is constructed. Operational indicators and emission indicators are integrated, and fuzzy inference rules are used for modeling.
The hierarchical relationship of cross-modal data was clarified, the interaction patterns of cross-equipment data and the inheritance characteristics of cross-stage data were explored, the data processing efficiency and the model's representation ability were improved, and the overall operating status of the desulfurization process could be comprehensively reflected.
Smart Images
Figure CN120910816A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of data processing, in particular to a data cross-modal representation and modeling method for a flue gas desulfurization process. BACKGROUND
[0002] The flue gas desulfurization process is a key link in industrial waste gas purification, and its system structure is complex, involving multiple hierarchical devices such as absorption towers, circulating pumps and oxidation blowers, and covering multiple process stages such as absorption, oxidation and product processing. A large amount of cross-modal data is generated in the desulfurization process, including device operation state data collected by sensors in real time, such as time series data such as temperature, pressure and flow, discrete log data such as device start and stop, process parameter configuration data and material composition detection data. These data sources are scattered and have certain hierarchical and dynamic correlations. At present, the cross-modal data of the flue gas desulfurization process is mostly modeled independently using single type data, resulting in chaotic attribution of cross-modal data, inability to provide structured input for modeling, and inability to effectively mine the interaction rules of cross-device data and the characteristics of cross-stage data. The lack of these dynamic correlation information makes it difficult for the model to reflect the overall running state of the desulfurization process.
[0003] Therefore, it is necessary to design a data cross-modal representation and modeling method for a flue gas desulfurization process to solve the problems existing in the prior art. SUMMARY
[0004] In view of this, the present application provides a data cross-modal representation and modeling method for a flue gas desulfurization process, aiming to solve the problems existing in the prior art.
[0005] The present application provides a data cross-modal representation and modeling method for a flue gas desulfurization process, comprising: Collecting cross-modal data of the flue gas desulfurization, the cross-modal data including multiple types of data of several hierarchical devices and several process stages of the desulfurization system; Processing the cross-modal data based on a hierarchical model and a time sequence correlation model, the hierarchical model being based on the desulfurization system to build hierarchical branches and determine hierarchical attribution relationships, and the time sequence correlation model taking a time sequence identifier as an index to determine node identifier records of the hierarchical model and generate cross-device correlation items and cross-stage correlation items; The deep autoencoder is constructed to perform feature extraction on the processed cross-modal data based on a multi-layer feature extraction network, and to determine a low-dimensional feature vector according to an extraction result; a decoder of the deep autoencoder is used to reconstruct the cross-modal data based on the low-dimensional feature vector, and to determine a reconstruction error between the reconstructed data and the cross-modal data, adjust the deep autoencoder based on the reconstruction error, and output a target low-dimensional feature vector according to the adjusted deep autoencoder. The operation index and the emission index of the flue gas desulfurization are collected, and a fuzzy neural network model is constructed based on the target low-dimensional feature vector, the operation index, the emission index, and a fuzzy inference rule.
[0006] Further, the multi-type data includes device operation log data, process parameter configuration data, process state data collected based on a sensor, and detected material property data.
[0007] Further, when the hierarchical model constructs a hierarchical branch based on the desulfurization system and determines a hierarchical attribution relationship, the following is included: The desulfurization system is taken as a root node, the desulfurization system is divided into a plurality of first-level child nodes according to a plurality of process stages, each first-level child node corresponds to a process stage, each first-level child node divides a plurality of hierarchical devices belonging thereto as second-level child nodes, each second-level child node corresponds to a single device, and each second-level child node divides a plurality of data collection points corresponding thereto as third-level child nodes. Based on the hierarchical association between the root node, each first-level child node, each second-level child node, and each third-level child node, a hierarchical attribution relationship of the cross-modal data is determined.
[0008] Further, when the time sequence association model determines a time sequence identifier as an index, determines a node identifier record of the hierarchical model, and generates a cross-device association item and a cross-stage association item, the following is included: The time sequence association model determines a timestamp of the time sequence identifier, uses the timestamp as an index, constructs a mapping relationship between the node identifier of each third-level child node of the hierarchical model, determines that the data of different single devices corresponding to a same timestamp generate the cross-device association item, and determines that the data of adjacent process stages generate the cross-stage association item, the cross-device association item and the cross-stage association item are used to represent the dynamic association between data.
[0009] Further, the deep autoencoder performs feature extraction on the processed cross-modal data based on a multi-layer feature extraction network, and determines a low-dimensional feature vector according to an extraction result, including: The multi-layer feature extraction network of the deep autoencoder includes a feature extraction layer and a fusion layer. The feature extraction layer extracts the time sequence features of the time sequence type data in the processed cross-modal data based on a convolution network, and extracts the discrete features of the discrete type data in the processed cross-modal data based on an embedding network, and uses an attention mechanism to strengthen the correlation features of the cross-device correlation items and the cross-stage correlation items, and the fusion layer performs dimension unification and information fusion on the strengthened correlation features, the time sequence features and the discrete features based on a full connection network, and outputs the low-dimensional feature vector.
[0010] Further, the deep autoencoder extracts features of the processed cross-modal data based on a multi-layer feature extraction network, and determines the low-dimensional feature vector according to the extraction result, and further comprises: The convolution network comprises a convolution layer and a pooling layer, the convolution layer is used to capture local features of the time sequence type data, and the pooling layer is used to reduce the feature dimension of the local features and determine the time sequence features.
[0011] Further, when the decoder of the deep autoencoder reconstructs the cross-modal data based on the low-dimensional feature vector and determines the reconstruction error between the reconstructed data and the cross-modal data, it comprises: The decoder of the deep autoencoder comprises a deconvolution layer and a restoration layer; The deconvolution layer corresponds to the inverse process of the convolution layer, and is used to reconstruct the time sequence features, the restoration layer corresponds to the inverse process of the embedding network, and is used to reconstruct the discrete features, and the reconstruction error is the difference between the reconstructed time sequence features and discrete features and the processed cross-modal data.
[0012] Further, when the deep autoencoder is adjusted based on the reconstruction error, and the target low-dimensional feature vector is output according to the adjusted deep autoencoder, it comprises: The network parameters of the deep autoencoder are adjusted based on the reconstruction error, and the deep autoencoder is re-extracted from the processed cross-modal data, and the target low-dimensional feature vector is output.
[0013] Further, when the fuzzy neural network model is constructed based on the target low-dimensional feature vector, the running index, the emission index and the fuzzy inference rule, it comprises: The fuzzy neural network model comprises an input layer, a fuzzification layer, a rule layer and an output layer; The input layer receives the target low-dimensional feature vector, the fuzzification layer converts the target low-dimensional feature vector into a fuzzy set based on a membership function, the rule layer performs rule inference based on the fuzzy inference rule, and the output layer outputs the performance index related to the running index and the emission index.
[0014] Further, in constructing the fuzzy neural network model based on the target low-dimensional feature vector, the operation index, the emission index and the fuzzy inference rule, further comprising: A sample data set is acquired, the membership function is determined according to the historical distribution of the sample data set, and the fuzzy inference rule is determined based on the association rule of the historical operation index and the historical emission index in the sample data set.
[0015] Compared with the prior art, the beneficial effects of the present application are that the hierarchical equipment and process stages of the flue gas desulfurization system are combed according to the hierarchical model, the hierarchical attribution relationship of the cross-modal data is clarified, the problem of chaotic data attribution in the past is effectively solved, and a structured input basis is provided for subsequent modeling. At the same time, the time sequence correlation model takes the time sequence identifier as the index, generates cross-equipment correlation items and cross-stage correlation items, mines the interaction rule of cross-equipment data and the carrying feature of cross-stage data, makes up for the short board of missing dynamic correlation information, the deep autoencoder extracts representative low-dimensional feature vectors through a multi-layer feature extraction network, reduces data redundancy, and improves data processing efficiency. The deep autoencoder is adjusted based on the reconstruction error, further improving the accuracy of feature extraction. The fuzzy neural network model constructed in combination with the target low-dimensional feature vector, the operation index, the emission index and the fuzzy inference rule can fully fuse multi-source information, so that the model can fully reflect the global operation state, thereby improving the representation ability of the model for the desulfurization process. BRIEF DESCRIPTION OF DRAWINGS
[0016] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The accompanying drawings are included to provide a description of preferred embodiments and are not meant to limit the present application. Moreover, the same reference numerals in different figures represent the same or similar components. In the drawings: Figure 1 A flowchart of a data cross-modal representation and modeling method for a flue gas desulfurization process according to an embodiment of the present application. DETAILED DESCRIPTION
[0017] Exemplary embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. Rather, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be accurately conveyed to those skilled in the art. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.
[0018] In some embodiments of the present application, referring to Figure 1 A modeling method for data cross-modal representation of a flue gas desulfurization process is shown, comprising: S100: Collecting cross-modal data of the flue gas desulfurization process, the cross-modal data containing multi-type data of several hierarchical devices and several process stages of the desulfurization system.
[0019] S200: Processing the cross-modal data based on a hierarchical model and a time sequence correlation model, the hierarchical model constructing hierarchical branches based on the desulfurization system and determining hierarchical attribution relationships, the time sequence correlation model taking time sequence identifiers as indexes, determining node identifier records of the hierarchical model, and generating cross-device correlation items and cross-stage correlation items.
[0020] S300: Constructing a deep autoencoder, the deep autoencoder performing feature extraction on the processed cross-modal data based on a multi-layer feature extraction network, determining a low-dimensional feature vector according to the extraction result, the decoder of the deep autoencoder reconstructing the cross-modal data based on the low-dimensional feature vector, determining reconstruction errors of the reconstructed data and the cross-modal data, adjusting the deep autoencoder based on the reconstruction errors, and outputting a target low-dimensional feature vector according to the adjusted deep autoencoder.
[0021] S400: Collecting operation indicators and emission indicators of the flue gas desulfurization process, and constructing a fuzzy neural network model based on the target low-dimensional feature vector, the operation indicators, the emission indicators, and fuzzy inference rules.
[0022] Specifically, the flue gas desulfurization system has many levels of equipment, such as absorption tower, circulating pump, etc., and many process stages, such as absorption process, oxidation process, etc. The data of the level equipment includes the real-time temperature of the absorption tower, the slurry liquid level height, the inlet and outlet flue gas pressure difference, etc. The data of the process stage includes the slurry pH value of the absorption process, the absorbent dosage, the oxidation air flow of the oxidation process, and the sulfite ion concentration, etc. Moreover, the data types generated by each process stage are relatively miscellaneous. Only by collecting these data comprehensively can the state of the flue gas desulfurization system be complete, and sufficient information can be provided for subsequent processing. The hierarchical model and the time sequence correlation model are used to process the cross-modal data. The hierarchical model is based on the desulfurization system to build a hierarchical branch, and clearly defines the hierarchical attribution of each data, thereby solving the problem of chaotic data attribution, so that each data can correspond to a specific equipment or process stage. The data between different devices and different process stages are associated, such as the circulating pump flow affecting the absorption tower pressure, and the product of the absorption link in the process stage affecting the efficiency of the oxidation link. They have dynamic correlations. Single data modeling cannot reflect these dynamic correlations, cannot effectively mine interactive rules and bearing characteristics, and the time sequence correlation model takes the time sequence identifier as the index, records the node identifier of the hierarchical model, generates cross-device correlation items and cross-stage correlation items, and captures the data correlation between different devices and different process stages. The cross-modal data processed by the hierarchical model and the time sequence correlation model is still complex, and the core features need to be extracted. A deep autoencoder of a multi-layer feature extraction network is constructed to obtain a low-dimensional feature vector. The multi-layer feature extraction network can peel off the redundant information in the data layer by layer, retain the key features, and generate a low-dimensional feature vector. The low-dimensional feature vector can simplify the data while retaining its core information. After being reconstructed by the decoder, the deep autoencoder is adjusted adaptively by using the reconstruction error to ensure that the output target low-dimensional feature vector can further retain key information and ensure that the extracted features are accurate, providing high-quality input for modeling.
[0023] It can be understood that the operation index is a key parameter reflecting the operation state of the flue gas desulfurization system equipment and the process efficiency, including the desulfurization efficiency of the absorption tower, the conveying efficiency of the circulating pump, the oxygen supply efficiency of the oxidation fan, etc. The operation index directly reflects whether the equipment is running normally and whether the process parameters are in the appropriate interval. The emission index focuses on the purification effect of the waste gas after desulfurization, including the concentration of sulfur dioxide, the concentration of smoke dust, the concentration of nitrogen oxides, the temperature and humidity of flue gas, and the purity and water content of the desulfurization byproduct. These indexes are the basis for measuring whether the desulfurization process meets the environmental protection standard. On the one hand, the operation index and the emission index are the direct embodiment of the desulfurization process. After the cross-modal data is processed and the features are extracted, it finally needs to be associated with the actual operation effect and the environmental protection standard. For example, the target low-dimensional feature vector contains equipment interaction rules and process stage characteristics. Only by combining the operation index and the emission index can the mapping relationship between these features and the desulfurization efficiency and environmental protection be determined. On the other hand, the fuzzy neural network model needs to combine the data features with the actual indexes to effectively evaluate and predict the desulfurization process. The establishment of fuzzy reasoning rules depends on the association rules between the operation index and the emission index and the data features. If the operation index and the emission index are missing, the model will not be able to link the data features with the actual operation effect, so it cannot accurately reflect the global state of the flue gas desulfurization process. Therefore, the fuzzy reasoning rule combines the target low-dimensional feature vector with the actual operation index and the emission index, so as to take into account the data rules, to make up for the lack of dynamic association information. Through hierarchical model and time series model processing, the structured association of cross-modal data can be effectively captured to solve the problems of attribution confusion and association loss. The target low-dimensional feature vector extracted by the deep autoencoder can preserve its key information, improve the effectiveness of feature extraction, provide high-quality input data for modeling, ensure the fitting degree and accuracy of the model to the data, and the fuzzy neural network combines the features with the operation index and the emission index to accurately model, providing a reliable basis for the desulfurization process management, improving the data utilization efficiency and the accuracy of modeling.
[0024] In some embodiments of the present application, the multi-type data includes: equipment operation log data, process parameter configuration data, process state data collected based on sensors, and detected material property data.
[0025] Specifically, the process state data collected by the sensors (such as the absorption tower temperature, the circulating pump flow, etc.), the equipment operation log data (such as the start-stop record, etc.), the process parameter configuration data (such as the pH value of the absorption process, the amount of absorbent, etc.), and the material characteristic data (such as the water content of gypsum) jointly constitute the specific carriers of the multi-type data. The equipment operation log data corresponds to the discrete logs of each equipment, the process parameter configuration data corresponds to the configuration data of each stage, the process state data collected by the sensors covers the time sequence data of the equipment and the process stage, and the detected material characteristic data corresponds to the material data of the product processing stage. These data cover multiple dimensions from equipment to process, from real-time state to operation benchmark, and from discrete events to continuous changes, and completely present the operation of the flue gas desulfurization system. The limitation of a single type of data is particularly obvious in cross-modal data. For example, only looking at the real-time temperature of the absorption tower (process state data collected by the sensors) cannot determine whether the temperature fluctuates in operation in combination with the start-stop record (equipment operation log data). The multi-type characteristics of the cross-modal data determine that the data must be effectively integrated to break the fragmentation of the data, so as to capture the interaction rules between various equipment and various process stages, provide comprehensive and related inputs for modeling, enable the model to fully reflect the overall operation state, and thus improve the modeling ability of the model to the desulfurization process.
[0026] In some embodiments of the present application, when the hierarchical model constructs the hierarchical branches based on the desulfurization system and determines the hierarchical attribution relationship, the desulfurization system is taken as a root node, the desulfurization system is divided into a plurality of first-level child nodes according to a plurality of process stages, each first-level child node corresponds to a process stage, each first-level child node divides a plurality of hierarchical equipment belonging to the first-level child node as second-level child nodes, each second-level child node corresponds to a single equipment, and each second-level child node divides a plurality of data collection points corresponding to the second-level child node as third-level child nodes. Based on the hierarchical association between the root node, each first-level child node, each second-level child node, and each third-level child node, the hierarchical attribution relationship of the cross-modal data is determined.
[0027] Specifically, taking the desulfurization system as the root node, the first-level child nodes are divided according to the absorption, oxidation and other process stages, and under each first-level child node, the devices used, such as the absorption tower, the circulating pump and the like, are taken as the second-level child nodes, and the data acquisition points of the devices, such as the arrangement position of the temperature sensor of the absorption tower and the position for monitoring the flow of the circulating pump, are taken as the third-level child nodes. Through the hierarchical association between nodes, the root node governs the first-level child nodes, the first-level child nodes are associated with the corresponding second-level child nodes, and the second-level child nodes are connected with the third-level child nodes, so that the cross-modal data can be clearly attributed to a level. If the attribution is not clear, there may be data confusion, for example, the absorption tower temperature data of the absorption process stage and the oxidation fan temperature data of the oxidation process stage may be incorrectly attributed to the same type of temperature data for processing. However, in fact, the temperature change of the absorption tower is related to the reaction rate of the absorbent, and the temperature change of the oxidation fan is related to the equipment heat dissipation state, and the influencing factors and process states of the two are completely different. Through the division of nodes, the system-process-equipment-acquisition point logic can be sorted out, and through clear hierarchical branches, the attribution of cross-modal data is clear, avoiding confusion. Moreover, the hierarchical association reflects the internal relationship between process and equipment, equipment and acquisition point, lays a foundation for mining cross-equipment and cross-stage data association, thereby solving the problem of data attribution confusion, providing accurate structured input for modeling, accurately reflecting the overall running state of the desulfurization process, and improving the reliability of the model.
[0028] In some embodiments of the present application, when the time sequence association model takes the time sequence identifier as an index, the node identifier record of the hierarchical model is determined, and the cross-equipment association item and the cross-stage association item are generated, comprising: the time sequence association model determines the time stamp of the time sequence identifier, takes the time stamp as an index, constructs a mapping relationship of the node identifier of each third-level child node of the hierarchical model, determines the data corresponding to different single devices at the same time stamp to generate a cross-equipment association item, and determines the data of adjacent process stages to generate a cross-stage association item. The cross-equipment association item and the cross-stage association item are used to represent the dynamic association between data.
[0029] Specifically, the cross-modal data covers different sources and different types of data in the desulfurization process, which are not isolated but attached to the hierarchical structure of the desulfurization system. By setting system-process-equipment-collection point and other hierarchical nodes, each type of cross-modal data can be mapped to a specific node of the hierarchical model. The time sequence correlation model uses time stamp-based time sequence identification to realize the time sequence anchoring of the data by constructing the mapping relationship between the time stamp and the node identification of the three-level sub-nodes in the hierarchical model. The node identification is a unique identifier assigned to each node in the hierarchical model, which is used to accurately locate the source and ownership of the data. The accurate node positioning provides a basis for generating cross-device association items and cross-stage association items, ensuring that the ownership of the dynamically associated data is clear and traceable. The time sequence identification establishes a time dimension on the node identification. For cross-device association items, the three-level sub-node data corresponding to different two-level sub-nodes (such as the absorption tower and the circulating pump) at the same time stamp will be associated, for example, the absorption tower liquid level data and the circulating pump flow data at a certain moment form an association item, reflecting the real-time interaction between devices. Adjacent process stages refer to adjacent preceding and subsequent process stages. For cross-stage association items, the data of the sub-nodes in the previous process stage (such as the absorption stage) and the corresponding data in the subsequent process stage (such as the oxidation stage) are associated according to adjacent time stamps. Device operation and process advancement have strong time dependence. For example, changes in circulating pump flow will immediately affect absorption tower pressure, and the desulfurization effect in the absorption stage directly determines the reaction starting point in the oxidation stage. These dynamic associations will span devices and process stages over time. If not indexed and associated by time stamp, relying solely on static hierarchical ownership will break these real-time and transfer relationships, making it difficult for the model to capture key interaction patterns and reflect the overall process. It can be understood that the accurate association of cross-device and cross-stage association items through time stamp, the clear dynamic association relationship, make up for the information loss of single modeling, and make the fragmented data form a whole association, providing associated and structured input for subsequent feature extraction. The integration of dynamic association information enables the model to truly restore the time sequence evolution and global linkage of the desulfurization process, improving the representation accuracy of the operating state.
[0030] In some embodiments of the present application, the deep autoencoder extracts features from the processed cross-modal data based on a multi-layer feature extraction network, and determines a low-dimensional feature vector according to the extraction result, including: the multi-layer feature extraction network of the deep autoencoder includes a feature extraction layer and a fusion layer, the feature extraction layer extracts time sequence features of time sequence type data in the processed cross-modal data based on a convolutional network, and extracts discrete features of discrete type data in the processed cross-modal data based on an embedding network, and uses an attention mechanism to strengthen the association features of the cross-device association items and the cross-stage association items, the fusion layer unifies the dimensions and fuses the information of the strengthened association features, the time sequence features and the discrete features based on a fully connected network, and outputs a low-dimensional feature vector.
[0031] Specifically, the multi-layer feature extraction network realizes feature fusion of the processed cross-modal data through hierarchical processing. After processing of the cross-modal data by the hierarchical model and the time sequence correlation model, the processed cross-modal data not only contains cross-device correlation items and cross-stage correlation items, but also includes cross-modal data determined by the hierarchical model to have a hierarchical attribution relationship (such as process state data, device log and other basic data according to system-process-device-collected point), the cross-device correlation items and the cross-stage correlation items are derived features generated by the time sequence correlation model based on the original data, and the two together constitute the processed cross-modal data. After processing, the types of the processed cross-modal data still have certain differences, the continuity of time sequence data and the discontinuity of discrete data cannot be directly fused, and the cross-device and cross-stage correlation features are the key to reflect the global state, and if not strengthened, they will be ignored. A single feature extraction method cannot adapt to multiple types of data, and the un-fused features have high dimension and are redundant, which increases the modeling complexity. In the feature extraction layer, the convolutional network is good at capturing the dynamic change rule of time sequence data (such as temperature and flow) to extract time sequence features, and the embedding network extracts discrete features of discrete data (such as start-stop log and fault alarm). At the same time, the attention mechanism strengthens the correlation features in the cross-device correlation items and the cross-stage correlation items. The fusion layer unifies the dimensions of the time sequence features, the discrete features and the strengthened correlation features through a full connection network, and integrates them into a low-dimensional feature vector for retaining the core information.
[0032] It can be understood that the convolutional network and the embedding network are respectively adapted to time sequence and discrete data, which improves the pertinence of feature extraction, and the attention mechanism strengthens the correlation features, highlights the dynamic correlation of cross-device and cross-stage, the generated low-dimensional feature vector not only reduces redundancy but also retains core information, reduces the complexity of the subsequent model, and at the same time improves the representation ability of the features, laying a foundation for fuzzy neural network modeling.
[0033] In some embodiments of the present application, the deep autoencoder extracts features from the processed cross-modal data based on the multi-layer feature extraction network, and determines a low-dimensional feature vector according to the extraction result, further comprising: the convolutional network includes a convolutional layer and a pooling layer, the convolutional layer is used to capture local features of time sequence data, and the pooling layer is used to reduce the feature dimension of the local features and determine time sequence features.
[0034] Specifically, the convolutional network extracts the core features of the time series data through the synergy of the convolutional layer and the pooling layer. The convolutional layer uses a sliding window convolution kernel to locally scan the time series data (such as temperature, flow rate changes over time) and capture the correlation patterns of adjacent time steps (such as temperature rising trend in 5 minutes, periodic fluctuations of flow rate), generating feature maps containing local details. The pooling layer compresses the local features by maximum or average sampling, reducing the feature dimension while retaining key information (such as temperature peak, flow average), and finally outputs the time series features. The synergy of convolutional layer and pooling layer is used to extract time series features of time series data because time series data has the characteristics of continuity and contains a large amount of redundant information. Direct processing of time series data will result in high feature dimension and large noise interference. For example, the temperature data of the absorption tower is collected every second, and the time series data of one day contains 86400 points, most of which are redundant information of smooth fluctuations. The local feature capture capability of the convolutional layer can focus on meaningful change patterns (such as temperature fluctuations), and the dimension reduction effect of the pooling layer can eliminate redundancy and reduce computational load, avoiding overfitting in subsequent modeling. If the collaboration of the two layers is lacking, the noise and redundancy of the original time series data will mask the key features, affecting the effectiveness of feature fusion. It can be understood that the convolutional layer can accurately capture the local dynamic features of the time series data, improve the recognition of the features, and the pooling layer effectively reduces the feature dimension, reduces the consumption of computing resources, and enhances the robustness of the features to noise. The time series features generated by the combination of the two layers not only retain the key dynamic information, but also ensure the reliability of the fusion of discrete features and associated features, providing high-quality time series input for the generation of low-dimensional feature vectors, and further improving the feature representation ability of the deep autoencoder. In some embodiments of the present application, when the decoder of the deep autoencoder reconstructs the cross-modal data based on the low-dimensional feature vector and determines the reconstruction error between the reconstructed data and the cross-modal data, the decoder of the deep autoencoder includes a deconvolutional layer and a restoration layer, the deconvolutional layer corresponds to the inverse process of the convolutional layer, the deconvolutional layer is used to reconstruct the time series features, the restoration layer corresponds to the inverse process of the embedding network, the restoration layer is used to reconstruct the discrete features, and the reconstruction error is the difference between the reconstructed time series features and discrete features and the processed cross-modal data.
[0035] Specifically, the role of the deep autoencoder is to extract features from the processed cross-modal data and achieve alignment of data standardization. As a neural network model, the standard structure of the deep autoencoder includes an encoder and a decoder. The encoder is responsible for mapping the input data to a low-dimensional feature vector (based on feature extraction of a multi-layer feature extraction network), while the decoder is responsible for reconstructing the low-dimensional feature vector into an output consistent with the format of the input data. The cross-modal data is reconstructed through the inverse process of the deconvolution layer and the restoration layer to achieve a mapping closed loop of features and original data. The deconvolution layer corresponds to the inverse operation of the convolution layer, which expands the feature dimension and restores the local feature distribution to reconstruct the time series features in the low-dimensional feature vector into time series data close to the original. The restoration layer corresponds to the inverse process of the embedding network, which converts discrete features from vector form to original discrete data. The reconstruction error is measured by calculating the difference (mean square error) between the reconstructed time series features, discrete features and the processed cross-modal data, which reflects the accuracy of feature extraction. The deep autoencoder verifies the effectiveness of the features through the closed loop of encoding-decoding. If only the features are extracted without reconstruction, it is impossible to determine whether the low-dimensional feature vector retains the core information of the original data. The deconvolution layer adapts to the continuity reconstruction of time series data, and the restoration layer adapts to the non-continuity restoration of discrete data. Through differential reconstruction, the error of different types of data can be accurately calculated.
[0036] It can be understood that the inverse processes of the deconvolution layer and the restoration layer match the characteristics of time series and discrete data respectively, which improves the accuracy of reconstruction. The determined reconstruction error can quantify the completeness of feature extraction, provide a clear direction for optimizing the deep autoencoder, and ensure that key information can be retained, so that modeling is based on accurate feature input, which improves the representation ability of the global state of the desulfurization process. In some embodiments of the present application, when the deep autoencoder is adjusted based on the reconstruction error, and the target low-dimensional feature vector is output according to the adjusted deep autoencoder, it includes: adjusting the network parameters of the deep autoencoder based on the reconstruction error, and re-extracting features from the processed cross-modal data based on the adjusted deep autoencoder to output the target low-dimensional feature vector.
[0037] Specifically, the reconstruction error is taken as the feedback signal, and the network parameters of the deep autoencoder, such as the convolution kernel weights of the convolution layer, the mapping matrix of the embedding network, and the connection weights of the fully connected layer, are adjusted through the back propagation algorithm. When the reconstruction error (mean square error) is large, it indicates that the feature extraction process has lost key information or introduced redundancy. At this time, the parameters are corrected through gradient descent to reduce the difference between the reconstructed data and the original cross-modal data. After multiple rounds of iterative adjustment, when the error reduction amplitude continues to decrease and the error change of continuous multiple iterations is less than 0.001, it indicates that the correction of the network parameters cannot significantly reduce the error, and the accuracy of feature extraction and data reconstruction enters a stable stage. The stable stage avoids insufficient feature extraction caused by insufficient parameters and prevents overfitting caused by excessive optimization of parameters, providing a guarantee for the output of the target low-dimensional feature vector. Moreover, the output target low-dimensional feature vector can accurately preserve the core information and dynamic correlation of the original cross-modal data, improving the reliability of feature extraction. In some embodiments of the present application, when constructing the fuzzy neural network model based on the target low-dimensional feature vector, the operation index, the emission index, and the fuzzy inference rule, the fuzzy neural network model includes an input layer, a fuzzification layer, a rule layer, and an output layer. The input layer receives the target low-dimensional feature vector, the fuzzification layer converts the target low-dimensional feature vector into a fuzzy set based on the membership function, the rule layer performs rule inference based on the fuzzy inference rule, and the output layer outputs the performance index related to the operation index and the emission index.
[0038] Specifically, the fuzzy neural network model fuses the target low-dimensional feature vector through a multi-layer structure, truly restores the time evolution and global linkage of the desulfurization process, and improves the representation accuracy of the operating state. The input layer is used to receive the target low-dimensional feature vector, and the fuzzification layer uses the membership function (such as the Gaussian function) to convert the feature vector into a fuzzy set, solving the problem of data uncertainty. The rule layer performs inference on the fuzzy set according to the fuzzy inference rule, and the output layer de-fuzzifies the inference result to output the performance index related to the operation index and the emission index, such as energy efficiency, environmental protection compliance rate, and process stability. Due to the strong nonlinearity and uncertainty of the desulfurization process, traditional modeling lacks the ability to handle uncertainty, and pure fuzzy rules are difficult to utilize data features. The fusion of the two can take advantage of both data-driven and knowledge-driven. The target low-dimensional feature vector has extracted core information, and can be efficiently associated with performance indicators in combination with fuzzy rules, avoiding the complexity of direct modeling of raw data, thereby improving the accuracy of modeling and the representation ability of the desulfurization process.
[0039] In some embodiments of the present application, when the fuzzy neural network model is constructed based on the target low-dimensional feature vector, the operation index, the emission index and the fuzzy inference rule, it further includes: obtaining a sample data set, the membership function is determined according to the historical distribution of the sample data set, and the fuzzy inference rule is determined based on the association rule of the historical operation index and the historical emission index in the sample data set.
[0040] Specifically, the sample data set includes a historical target low-dimensional feature vector, which is a core feature extracted by a deep autoencoder after processing historical cross-modal data, and covers key information such as historical device levels and historical process stage data. These data records the running conditions of the desulfurization system at different historical operating conditions and different time periods, and contains the association between various types of data, providing actual basis for determining the membership function and refining the fuzzy inference rule. Through the sample data set, data support is provided for the establishment of the fuzzy neural network model. The determination of the membership function is based on the historical distribution in the sample data, for example, by analyzing the fluctuation range of historical temperature data, the fuzzy interval boundaries of "low temperature", "medium temperature" and "high temperature" are set, so that the fuzzification process conforms to the actual data rule. The fuzzy inference rule is refined from the association rule of the sample data, such as the historical data showing that "when the absorption tower temperature is higher than 65℃ and the circulating pump flow is lower than 200m³ / h, the sulfur dioxide emission concentration is easy to exceed 35mg / m³", which is converted into an executable inference rule. Based on this, the model can better fit the actual characteristics of the desulfurization process, avoid errors caused by subjective speculation, and enable the fuzzy neural network model to accurately map the relationship between input features and performance indicators, so that the model can fully reflect the overall running state, thereby improving the model's representation ability for the desulfurization process.
[0041] In summary, the beneficial effects of the present application are as follows: according to the hierarchical model, the hierarchical equipment and process stages of the flue gas desulfurization system are sorted out, the hierarchical attribution relationship of the cross-modal data is clarified, the problem of chaotic data attribution in the past is effectively solved, and a structured input basis is provided for subsequent modeling. At the same time, the time sequence association model uses time sequence identifier as index to generate cross-device association items and cross-stage association items, and excavates the interaction rule of cross-device data and the transfer characteristics of cross-stage data, making up for the short board of missing dynamic association information. The deep autoencoder extracts representative low-dimensional feature vectors through a multi-layer feature extraction network to process the cross-modal data, reduces data redundancy, and improves data processing efficiency. The deep autoencoder is adjusted based on the reconstruction error to further improve the accuracy of feature extraction. The fuzzy neural network model constructed based on the target low-dimensional feature vector, the operation index, the emission index and the fuzzy inference rule can fully integrate multi-source information, so that the model can fully reflect the overall running state, thereby improving the model's representation ability for the desulfurization process.
[0042] Those skilled in the art will appreciate that embodiments of the application can be supplied as a method, a system or a computer program product. Accordingly, the application can be embodied in the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the application can be embodied in the form of a computer program product on one or more computer readable storage media (including, without limitation, magnetic disks; optical disks; magneto-optical disks; ROMs; flash memory; etc.) having computer usable program code embodied therein.
[0043] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.
[0044] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.
[0045] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.
[0046] Finally, it should be noted that the above-mentioned embodiments are merely intended for describing and illustrating, not limiting, the technical solutions of the present application. Although the present application has been described in detail with reference to the above-mentioned embodiments, those skilled in the art should understand that the specific embodiments of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application, and any modification or equivalent replacement without departing from the spirit and scope of the present application should be covered in the protection scope of the claims of the present application.
Claims
1. A modeling method for data cross-modal representation of a flue gas desulfurization process, characterized in that, The method comprises the following steps: Collecting cross-modal data of flue gas desulfurization, the cross-modal data including multi-type data of several levels of equipment and several process stages of the desulfurization system; Processing the cross-modal data based on a hierarchical model and a time sequence correlation model, the hierarchical model constructing hierarchical branches based on the desulfurization system and determining hierarchical attribution relationships, the time sequence correlation model taking time sequence identifiers as indexes, determining node identifier records of the hierarchical model, and generating cross-equipment correlation items and cross-stage correlation items; Constructing a deep autoencoder, the deep autoencoder performing feature extraction on the processed cross-modal data based on a multi-layer feature extraction network, determining a low-dimensional feature vector according to the extraction result, the decoder of the deep autoencoder reconstructing the cross-modal data based on the low-dimensional feature vector, determining reconstruction errors of the reconstructed data and the cross-modal data, adjusting the deep autoencoder based on the reconstruction errors, and outputting a target low-dimensional feature vector according to the adjusted deep autoencoder; Collecting operation indexes and emission indexes of the flue gas desulfurization, and constructing a fuzzy neural network model based on the target low-dimensional feature vector, the operation indexes, the emission indexes, and fuzzy inference rules.
2. The modeling method of data cross-modal representation for flue gas desulfurization process according to claim 1, characterized in that, The multi-type data includes device operation log data, process parameter configuration data, process state data collected based on sensors, and material property data detected.
3. The modeling method of data cross-modal representation for flue gas desulfurization process according to claim 2, characterized in that, When the hierarchical model constructs hierarchical branches based on the desulfurization system and determines hierarchical attribution relationships, the method comprises the following steps: Taking the desulfurization system as a root node, dividing the desulfurization system into several first-level child nodes according to several process stages, each first-level child node corresponding to a process stage, dividing several level equipment belonging to each first-level child node as second-level child nodes, each second-level child node corresponding to a single device, and dividing several data collection points corresponding to each second-level child node as third-level child nodes; Determining the hierarchical attribution relationships of the cross-modal data based on the hierarchical correlation between the root node, each first-level child node, each second-level child node, and each third-level child node.
4. The modeling method of data cross-modal representation for flue gas desulfurization process according to claim 3, characterized in that, When the time sequence correlation model takes time sequence identifiers as indexes, determines node identifier records of the hierarchical model, and generates cross-equipment correlation items and cross-stage correlation items, the method comprises the following steps: The time sequence correlation model determines time stamps of the time sequence identifiers, takes the time stamps as indexes, constructs a mapping relationship between the time stamps and node identifiers of each third-level child node of the hierarchical model, determines that data of different single devices corresponding to the same time stamp generate the cross-equipment correlation items, and determines that data of adjacent process stages generate the cross-stage correlation items, the cross-equipment correlation items and the cross-stage correlation items being used to represent dynamic correlations between data.
5. The modeling method of data cross-modal representation for flue gas desulfurization process according to claim 4, characterized in that, The deep autoencoder performs feature extraction on the processed cross-modal data based on a multi-layer feature extraction network, and determines a low-dimensional feature vector according to the extraction result, the method comprising the following steps: The multi-layer feature extraction network of the deep autoencoder includes feature extraction layers and fusion layers. The feature extraction layer extracts the time sequence feature of the time sequence type data in the processed cross-modal data based on a convolution network, and extracts the discrete feature of the discrete type data in the processed cross-modal data based on an embedding network, and uses an attention mechanism to strengthen the correlation feature of the cross-device correlation item and the cross-stage correlation item, and the fusion layer performs dimension unification and information fusion on the strengthened correlation feature, the time sequence feature and the discrete feature based on a full connection network, and outputs the low-dimensional feature vector.
6. The method of modeling data cross-modal representation for a flue gas desulfurization process according to claim 5, wherein, The deep autoencoder extracts features of the processed cross-modal data based on a multi-layer feature extraction network, and determines the low-dimensional feature vector according to the extraction result, and further comprises: The convolution network comprises a convolution layer and a pooling layer, the convolution layer is used to capture local features of the time sequence type data, and the pooling layer is used to reduce the feature dimension of the local features and determine the time sequence feature.
7. The modeling method of data cross-modal representation for flue gas desulfurization process according to claim 6, characterized in that, When the decoder of the deep autoencoder reconstructs the cross-modal data based on the low-dimensional feature vector and determines the reconstruction error between the reconstructed data and the cross-modal data, it comprises: The decoder of the deep autoencoder comprises a deconvolution layer and a restoration layer; The deconvolution layer corresponds to the inverse process of the convolution layer, and is used to reconstruct the time sequence feature, the restoration layer corresponds to the inverse process of the embedding network, and is used to reconstruct the discrete feature, and the reconstruction error is the difference between the reconstructed time sequence feature and discrete feature and the processed cross-modal data.
8. The modeling method of data cross-modal representation for flue gas desulfurization process according to claim 7, characterized in that, When the deep autoencoder is adjusted based on the reconstruction error, and the target low-dimensional feature vector is output according to the adjusted deep autoencoder, it comprises: The network parameters of the deep autoencoder are adjusted based on the reconstruction error, and the deep autoencoder is re-extracted from the processed cross-modal data to output the target low-dimensional feature vector.
9. The modeling method of data cross-modal representation for flue gas desulfurization process according to claim 8, characterized in that, When the fuzzy neural network model is constructed based on the target low-dimensional feature vector, the running index, the emission index and the fuzzy inference rule, it comprises: The fuzzy neural network model comprises an input layer, a fuzzification layer, a rule layer and an output layer; The input layer receives the target low-dimensional feature vector, the fuzzification layer converts the target low-dimensional feature vector into a fuzzy set based on a membership function, the rule layer performs rule inference based on the fuzzy inference rule, and the output layer outputs the performance index related to the running index and the emission index.
10. The modeling method of data cross-modal representation for flue gas desulfurization process according to claim 9, characterized in that, When the fuzzy neural network model is constructed based on the target low-dimensional feature vector, the running index, the emission index and the fuzzy inference rule, it further comprises: A sample data set is obtained, the membership function is determined according to the historical distribution of the sample data set, and the fuzzy inference rule is determined based on the correlation rule of the historical running index and the historical emission index in the sample data set.