Intelligent operation monitoring system and method for petrochemical production device
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- LANGFANG DEVELOPMENT ZONE CNPC ZHONGZHOU ENGINEERING SUPERVISION CO LTD
- Filing Date
- 2026-03-10
- Publication Date
- 2026-06-12
Smart Images
Figure CN122196631A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of intelligent monitoring technology, and more specifically, to an intelligent operation monitoring system and method for petrochemical production facilities. Background Technology
[0002] Hydrocracking units are core secondary processing units in petrochemical production, and their operating status directly affects the economic benefits and production safety of the entire refining and chemical enterprise. In actual production, hydrocracking units often exhibit multiple stable operating modes due to feedstock switching, load adjustment, or changes in product schemes. During the switching of operating modes, there are state fluctuations with drastic parameter fluctuations. If these fluctuations are not identified and effectively controlled in a timely manner, they can easily lead to safety accidents such as catalyst coking, reactor overheating, or even unplanned shutdowns.
[0003] In existing technologies, the operation monitoring of hydrocracking units typically employs single-parameter threshold alarms or simple pattern recognition methods based on historical data. However, these methods require interpolation preprocessing to handle irregular data with inconsistent sampling frequencies, asynchronous timing, and random missing data, which distorts the irregular patterns in the original data or introduces artificial noise. Furthermore, existing methods often process each sensor variable independently, making it difficult to capture the collaborative changes of different variables during transition states. They also lack sensitivity to early trends during operating condition switching, failing to achieve early warning and proactive intervention in transition states. Therefore, how to accurately identify and provide early warnings of multimodal operating conditions based on the irregular characteristics of the original data and the collaborative changes of different variables has become a challenge for the industry. Summary of the Invention
[0004] This application provides an intelligent operation monitoring system and method for petrochemical production plants, which can accurately identify and provide early warning of multimodal operating conditions during operation based on the irregular characteristics of raw data and the coordinated changes of different variables.
[0005] In a first aspect, this application provides an intelligent operation monitoring method for a petrochemical production plant, comprising the following steps:
[0006] Collect multi-source sensor data during the operation of hydrocracking units in petrochemical production;
[0007] All numerical variables, data missing masks, and sampling time intervals in the multi-source sensor data are encoded using multi-channel image encoding to construct a current operating condition diagram of the hydrocracking unit.
[0008] The operating condition diagram is divided into blocks for mapping to obtain the operating condition association network of the multi-modal operating conditions of the hydrocracking unit. The multi-scale relative difference feature vector of the operating condition association network is constructed to obtain the enhanced node features under the current operating condition mode.
[0009] Acquire historical operating data of the hydrocracking unit, and determine multiple stable operating condition diagrams based on the historical operating data to identify the operating conditions;
[0010] Multiple matching degrees are determined for all stable operating condition maps and all node features. When all matching degrees are lower than a preset threshold, the operating mode of the hydrocracking unit is determined to be a transition state. A blue warning is issued at the beginning stage, and the control parameters are adjusted at the evolution stage to suppress the operating condition fluctuations of the transition state.
[0011] In some embodiments, constructing a current operating condition diagram of the hydrocracking unit by performing multi-channel image encoding on all numerical variables, data missing masks, and sampling time intervals in the multi-source sensor data specifically includes:
[0012] Time alignment is performed on the multi-source sensor data;
[0013] Extract the numerical variables of each sampling time from the aligned multi-source sensor data, obtain the numerical matrix, and input it into the numerical channel;
[0014] Construct a missing indicator matrix based on the missing distribution of the numerical matrix data and input the missing channels;
[0015] Determine the sampling interval matrix for the time interval of data acquisition and input the forward hold interval channel;
[0016] By aligning the numerical channels, missing channels, and interval channels, a current operating status diagram of the hydrocracking unit is obtained.
[0017] In some embodiments, the process of segmenting and mapping the operating condition diagram to obtain the operating condition association network of the multimodal operating conditions of the hydrocracking unit specifically includes:
[0018] The working condition diagram is evenly divided into multiple image blocks, and all image blocks are converted into initial node feature vectors through linear mapping;
[0019] Obtain the grid coordinates of each image block in the working condition diagram;
[0020] Using the node corresponding to each image patch as the center, the final connected neighbor nodes of each image patch are determined based on the feature distance between the initial node feature vectors of each node.
[0021] By constructing connecting edges with all nodes and their corresponding selected neighbor nodes, and then undirecting all edges, the operating condition association network of the hydrocracking unit is obtained.
[0022] In some embodiments, constructing multi-scale relative difference feature vectors of the working condition association network to obtain enhanced node features under the current working condition mode specifically includes:
[0023] Traverse each edge in the aforementioned working condition association network to obtain the characteristics of the central node and neighboring nodes of each edge;
[0024] Determine the directional difference features and amplitude difference features between the features of the neighboring nodes and the features of the center node.
[0025] By combining interactive coupling features, the corresponding multi-scale relative difference feature vectors are obtained;
[0026] Attention-weighted aggregation and residual connection are performed on all relative difference feature vectors to obtain enhanced node features under the current working condition mode.
[0027] In some embodiments, determining multiple stable operating condition diagrams based on the historical operating data specifically includes:
[0028] Historical sensor data of the hydrocracking unit under multiple known operating conditions are acquired, and multi-channel image encoding is performed on all historical sensor data to obtain the corresponding historical operating condition state diagram.
[0029] Perform block mapping and network construction on the historical working condition state diagram to obtain the historical association network corresponding to the historical sample;
[0030] Multi-scale relative difference feature vector construction and node feature enhancement are performed on all historical correlation networks to obtain multiple stable working condition maps that identify the working condition status.
[0031] In some embodiments, determining multiple matching degrees between all stable condition atlases and all node features specifically includes:
[0032] Obtain the preset steady-state determination threshold;
[0033] Calculate the similarity between the features of each node and all stable operating condition maps, and use all similarities as the corresponding matching degree;
[0034] Compare all matching degrees with the steady-state determination threshold.
[0035] In some embodiments, issuing a blue warning during the initial stage and adjusting control parameters during the evolution stage to suppress transient operating condition fluctuations specifically includes:
[0036] Identify the initial and evolutionary stages of the transition state;
[0037] A blue warning is triggered during the initial phase, and the warning information is pushed to the central control station to remind operators to pay attention to changes in the current operating conditions.
[0038] During the evolution phase, the control parameters of the hydrocracking unit are adjusted based on the changing trends of all relative differential eigenvectors in the operating condition correlation network. These control parameters include reactor inlet temperature, circulating hydrogen flow rate, and high-pressure water injection rate.
[0039] Secondly, this application provides an intelligent operation monitoring system for petrochemical production plants, comprising:
[0040] The acquisition module is used to collect multi-source sensor data during the operation of hydrocracking units in petrochemical production.
[0041] The processing module is used to perform multi-channel image encoding on all numerical variables, data missing masks and sampling time intervals in the multi-source sensor data to construct the operating status diagram of the hydrocracking unit during current operation.
[0042] The processing module is also used to perform block mapping on the operating condition diagram to obtain the operating condition association network of the multi-modal operating conditions of the hydrocracking unit, construct the multi-scale relative difference feature vector of the operating condition association network, and then obtain the enhanced multiple node features under the current operating condition mode.
[0043] The processing module is also used to acquire historical operating data of the hydrocracking unit and determine multiple stable operating condition diagrams based on the historical operating data to identify the operating conditions.
[0044] The execution module is used to determine multiple matching degrees of all stable operating condition maps and all node features. When all matching degrees are lower than a preset threshold, the operating mode of the hydrocracking unit is determined to be a transition state. A blue warning is issued at the beginning stage, and the control parameters are adjusted at the evolution stage to suppress the operating condition fluctuations of the transition state.
[0045] Thirdly, this application provides a computer device, the computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described intelligent operation monitoring method for petrochemical production equipment.
[0046] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described intelligent operation monitoring method for petrochemical production equipment.
[0047] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects:
[0048] This application provides an intelligent operation monitoring system and method for petrochemical production units. First, multi-source sensor data of the hydrocracking unit during petrochemical production is collected. All numerical variables, missing data masks, and sampling time intervals in the multi-source sensor data are encoded using multi-channel image encoding to construct a current operating condition map of the hydrocracking unit. The operating condition map is then segmented and mapped to obtain a multi-modal operating condition correlation network for the hydrocracking unit. A multi-scale relative difference feature vector is constructed from this network to obtain enhanced node features under the current operating condition mode. Historical operating data of the hydrocracking unit is acquired, and multiple stable operating condition maps for identifying the operating condition are determined based on this data. Multiple matching degrees between all stable operating condition maps and all node features are determined. When all matching degrees are below a preset threshold, the operating mode of the hydrocracking unit is determined to be a transitional state, and a blue warning is issued at the initial stage. During the evolution stage, control parameters are adjusted to suppress operating condition fluctuations in the transitional state.
[0049] Therefore, in the intelligent operation monitoring method for a petrochemical production unit, this application firstly collects multi-source sensor data during the operation of the hydrocracking unit in petrochemical production; then, it performs multi-channel image encoding on all numerical variables, data missing masks, and sampling time intervals in the multi-source sensor data to construct a working condition status diagram of the hydrocracking unit at its current operation; wherein, the working condition status diagram is a three-channel image data that comprehensively reflects the sensor values, data missing status, and sampling interval information at the current operating moment of the hydrocracking unit, used to transform the originally irregular, multi-source, and asynchronous raw sensor data into a structured image representation, facilitating subsequent block processing and graph structure construction, while avoiding noise and artifacts introduced by traditional interpolation methods, and fully preserving the numerical dynamics, missing patterns, and non-uniform sampling characteristics in the original data; secondly, The operating condition diagram is segmented and mapped to obtain an operating condition association network for the multimodal operating conditions of the hydrocracking unit. Multi-scale relative difference feature vectors of this network are constructed to obtain enhanced node features for the current operating condition mode. These enhanced node features include both node-specific information and multi-scale difference information from neighboring nodes, enabling more sensitive perception of operating condition changes. Historical operating data of the hydrocracking unit is acquired, and multiple stable operating condition diagrams are determined based on this data to identify the operating condition state. Multiple matching degrees between all stable operating condition diagrams and all node features are determined. When all matching degrees are below a preset threshold, the operating mode of the hydrocracking unit is determined to be a transitional state, and a blue warning is issued at the initial stage. During the evolution stage, control parameters are adjusted to suppress operating condition fluctuations in the transitional state. This scheme can achieve accurate identification and early warning of multimodal operating conditions during operation based on the irregular characteristics of the original data and the coordinated changes of different variables. Attached Figure Description
[0050] Figure 1 This is an exemplary flowchart of an intelligent operation monitoring method for petrochemical production plants according to some embodiments of this application;
[0051] Figure 2 This is an exemplary flowchart illustrating the determination of operating condition states according to some embodiments of this application;
[0052] Figure 3 This is an exemplary flowchart illustrating the determination of a steady-state operating condition diagram according to some embodiments of this application;
[0053] Figure 4 This is a schematic diagram of the structure of an intelligent operation system for a petrochemical production unit according to some embodiments of this application;
[0054] Figure 5 This is a schematic diagram of the structure of a computer device for implementing an intelligent operation monitoring method for petrochemical production plants, according to some embodiments of this application. Detailed Implementation
[0055] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.
[0056] refer to Figure 1 The figure is an exemplary flowchart of an intelligent operation monitoring method for a petrochemical production plant according to some embodiments of this application. The intelligent operation monitoring method for the petrochemical production plant mainly includes the following steps:
[0057] In step 101, multi-source sensor data are collected during the operation of the hydrocracking unit in petrochemical production.
[0058] In specific implementation, the collection of multi-source sensor data during the operation of a hydrocracking unit in petrochemical production can be achieved in the following way: First, deploy various types of sensors in key parts of the hydrocracking unit; key parts include reactors, circulating hydrogen compressors, high-pressure heat exchangers, hydrogen feed pumps, and product fractionation towers; the deployed sensor types include at least vibration sensors, shaft displacement sensors, bearing temperature sensors, lubricating oil pressure sensors, lubricating oil temperature sensors, process medium flow sensors, process medium pressure sensors, and process medium temperature sensors; all sensors continuously collect data according to a preset sampling frequency, which is set according to the sensor type and monitoring requirements, for example, 500Hz. Each sensor records the current timestamp while collecting data, forming the original sampling time point; the original data collected by each sensor is transmitted to a data acquisition server, which performs preliminary processing of the received data, organizing it into original data records according to sensor number and time sequence; each original data record contains at least three fields: sensor number, sampling time point, and measured value, thereby obtaining multi-source sensor data of the hydrocracking unit during petrochemical production; other embodiments may also use other methods, which are not limited here.
[0059] It should be noted that the multi-source sensing data in this application refers to a collection of various measurement data that reflect the operating status and process parameters of the equipment, collected from different parts and different types of sensors in the device; including vibration data, shaft displacement data, bearing temperature data, lubricating oil pressure data, lubricating oil temperature data, process medium flow data, process medium pressure data, and process medium temperature data.
[0060] In step 102, all numerical variables, data missing masks, and sampling time intervals in the multi-source sensor data are encoded using multi-channel image encoding to construct a working condition diagram of the hydrocracking unit during its current operation.
[0061] In some embodiments, reference Figure 2 As shown, this figure is an exemplary flowchart for determining the operating condition state diagram in some embodiments of this application. In this embodiment, the operating condition state diagram of the hydrocracking unit under current operation can be constructed by multi-channel image encoding of all numerical variables, data missing masks, and sampling time intervals in the multi-source sensor data using the following steps:
[0062] First, in step 1021, the multi-source sensor data is time-aligned;
[0063] Secondly, in step 1022, the numerical variables of each sampling time in the aligned multi-source sensing data are extracted to obtain a numerical matrix and input into the numerical channel.
[0064] Then, in step 1023, a missing indicator matrix is constructed based on the missing distribution of the numerical matrix data and the missing channel is input;
[0065] Furthermore, in step 1024, the sampling interval matrix of the time interval for data acquisition is determined and the forward hold interval channel is input;
[0066] Finally, in step 1025, the numerical channel, missing channel, and interval channel are aligned to obtain the operating status diagram of the hydrocracking unit during current operation.
[0067] In specific implementation, time alignment of the multi-source sensor data can be achieved in the following way: using the time point with the highest sampling frequency among all sensors in the multi-source sensor data as the benchmark, or establishing a unified time axis with a fixed time interval, such as 1 second; for each sensor, at each moment on the unified time axis, if the sensor has an original sampling point, then the value is directly used; otherwise, a forward padding method is used, that is, the most recent valid sampling value before that moment is taken as the value of the current moment; after time alignment is completed, the aligned data sequence of each sensor is organized into a matrix form, with the rows of the matrix corresponding to each moment on the unified time axis and the columns of the matrix corresponding to each sensor variable, and the time-aligned multi-source sensor data is used as the input for subsequent processing; other embodiments may also use other methods to achieve this, which are not limited here.
[0068] In specific implementation, extracting the numerical variables of each sampling time in the aligned multi-source sensor data to obtain a numerical matrix and inputting it into the numerical channel can be achieved in the following way: First, extract the measured values of the actual operating parameter variables collected by all sensors from the aligned multi-source sensor data; second, for each sampling time and each parameter variable, fill the corresponding measured value of the parameter variable into the corresponding position of a two-dimensional matrix. The rows of this two-dimensional matrix correspond to each time point on a unified time axis, arranged from top to bottom in chronological order; the columns of this two-dimensional matrix correspond to the actual operating parameter variables collected by each sensor, arranged from left to right according to sensor type or installation location. The matrix constructed in this way is the numerical matrix, used to completely record the numerical changes of the parameter variables of all sensors in a continuous time series; finally, the above numerical matrix is used as the first channel of the three-channel image, which is called the numerical channel, used to store the original measurement information of each parameter variable at each time point. Each element in the numerical matrix reflects the actual operating state of the corresponding variable at the corresponding time point. The entire numerical matrix is used as the input data of the numerical channel; other embodiments can also be implemented in other ways, which are not limited here.
[0069] In specific implementation, constructing a missing indicator matrix based on the missing distribution of the numerical matrix data and inputting it into the missing channel can be achieved in the following way: During the time alignment process, if there is a measurement value collected at a certain moment, i.e., a variable, without an original sampled value and unable to obtain a valid value through forward padding, for example, there is no forward value at the beginning of the sequence, then the measurement value of the variable at that moment is considered as missing data; construct a missing indicator matrix with the same number of rows and columns as the numerical matrix, where the rows of the missing indicator matrix correspond to each moment on a unified time axis, and the columns of the missing indicator matrix correspond to the parameter variables of each sensor; for each position in the missing indicator matrix, if the variable has a valid measurement value at that moment, then mark it as 0 at the corresponding position in the missing indicator matrix; if the measurement value of the variable at that moment is missing, then mark it as 1 at the corresponding position in the missing indicator matrix, thereby obtaining the missing indicator matrix and inputting it into the second channel of the three-channel image, which is called the missing channel; other embodiments can also be implemented in other ways, which are not limited here.
[0070] It should be noted that the missing data indication matrix in this application can directly reflect the data missing patterns and distribution of each variable at each time point. It is used as the second channel of the three-channel image, which is called the missing channel. It is used to store information on whether each variable has a valid measurement value at each time point. For example, in a hydrocracking unit, a vibration sensor may lose data due to a momentary failure. The missing data indication matrix will mark the corresponding position as 1 in the column corresponding to the vibration sensor and the row corresponding to the time of failure. This allows the subsequent model to perceive this information and use the missing data indication matrix as input data for the missing data channel for subsequent channels to merge.
[0071] In specific implementation, determining the sampling interval matrix for the time interval of data acquisition and inputting it into the forward-held interval channel can be achieved in the following way: For the data acquired by each sensor, calculate the time difference between adjacent valid sampling points based on its original sampling timestamp; on a unified time axis, for each moment, if there is valid data at that moment, whether it is the original value or the forward-filled value, record the time interval between that moment and the previous valid data moment; if that moment is the first valid data point, the time interval can be set to 0 or a preset value, and the time interval values of all moments of all sensors are combined to form a sampling interval matrix of the same size as the numerical matrix; secondly, in order to eliminate the difference in dimensions between different sensors, the time interval can be normalized, for example, by dividing it by the maximum time interval, so that its value range is between 0 and 1, thereby obtaining the sampling interval matrix, and using the sampling interval matrix as the input data of the third channel of the three-channel image, that is, the interval channel; the interval channel can capture the irregularity of sensor sampling, such as the low sampling frequency of some variables or uneven sampling intervals, for accurate identification of working conditions, and the sampling interval matrix is used as the input of the interval channel; other embodiments can also be implemented in other ways, which are not limited here.
[0072] In specific implementation, aligning the numerical channel, missing channel, and interval channel to obtain the operating condition diagram of the hydrocracking unit during current operation can be achieved in the following way: using the numerical matrix as the R channel of an RGB image, the missing indicator matrix as the G channel of an RGB image, and the sampling interval matrix as the B channel of an RGB image; wherein, the above three channels have the same height and width, the height equal to the number of moments on a unified time axis, and the width equal to the number of sensor variables; stacking the data of these three channels according to the channel dimensions to form a three-dimensional array, thus obtaining an RGB format operating condition diagram of the hydrocracking unit during current operation; this operating condition diagram fully preserves the dynamic information, missing patterns, and non-uniform sampling characteristics of the original multi-source sensor data, without any interpolation processing, providing a true and unbiased input for subsequent operating condition identification; using the numerical matrix, missing indicator matrix, and sampling interval matrix as the three channels of an RGB image respectively, and merging them to obtain the operating condition diagram of the hydrocracking unit during current operation; other embodiments may also use other methods, which are not limited here.
[0073] It should be noted that the operating condition diagram in this application is a three-channel image data that comprehensively reflects the sensor values, data missingness, and sampling interval information of the hydrocracking unit at the current operating moment. It is used to transform the originally irregular, multi-source, and asynchronous raw sensor data into a structured image representation, facilitating subsequent block processing and graph structure construction. This avoids the noise and artifacts introduced by traditional interpolation methods, and fully preserves the numerical dynamics, missing patterns, and non-uniform sampling characteristics of the original data. Specifically, the multi-source sensor data of the hydrocracking unit exhibits irregularities such as inconsistent sampling frequencies, asynchronous time, and random data missingness. Traditional methods typically use interpolation, but interpolation introduces artificial noise and distorts the irregular patterns in the original data. This step first unifies all sensor data to the same time base through time alignment, resolving the differences in data from multiple sources. The system first addresses the problem by constructing a numerical matrix to store the actual measured values of each variable at each time point, thus preserving the original numerical dynamics. Next, a missing data indicator matrix is constructed to directly encode the missing data pattern, using a binary representation to indicate whether a valid measurement exists at each location. This allows subsequent models to perceive missing information rather than passively accepting interpolation results. Finally, a sampling interval matrix is constructed to characterize the non-uniformity of sampling from each sensor, making the irregularity of sampling itself a learnable feature. This process transforms one-dimensional irregular time-series data into two-dimensional structured images, avoiding noise and artifacts introduced by interpolation and fully preserving the numerical dynamics, missing patterns, and sampling features of the original data. By converting irregular data into a regular image structure, it lays the data foundation for subsequent image segmentation and graph structure construction, enabling the previously difficult-to-handle asynchronous sampling and random missing data problems to be solved within a unified framework.
[0074] In step 103, the operating condition diagram is divided into blocks for mapping to obtain the operating condition association network of the multi-modal operating conditions of the hydrocracking unit. The multi-scale relative difference feature vector of the operating condition association network is constructed to obtain the enhanced node features under the current operating condition mode.
[0075] In some embodiments, the operation condition diagram is divided into blocks for mapping to obtain the operation condition association network of the multimodal operation conditions of the hydrocracking unit. This can be achieved by the following steps:
[0076] The working condition diagram is evenly divided into multiple image blocks, and all image blocks are converted into initial node feature vectors through linear mapping;
[0077] Obtain the grid coordinates of each image block in the working condition diagram;
[0078] Using the node corresponding to each image patch as the center, the final connected neighbor nodes of each image patch are determined based on the feature distance between the initial node feature vectors of each node.
[0079] By constructing connecting edges with all nodes and their corresponding selected neighbor nodes, and then undirecting all edges, the operating condition association network of the hydrocracking unit is obtained.
[0080] In specific implementation, the working condition state diagram is uniformly divided into multiple image blocks. Converting all image blocks into initial node feature vectors through linear mapping can be achieved in the following way: the working condition state diagram is uniformly divided according to a preset block size; for example, the working condition state diagram can be divided into image blocks of size h×w, where h represents the height in the time axis direction, i.e., the number of time steps; w represents the width in the variable axis direction, i.e., the number of sensors; each image block covers a set of sensor variables over a continuous period; for each image block, all pixel values within it, including the R, G, and B channels, are flattened into a one-dimensional vector. Each image block is considered a node, and then a linear mapping layer, i.e., a fully connected layer, is used to convert the corresponding vector into a fixed-dimensional initial node feature vector; each image block corresponds to a graph node, and the vector obtained by linear mapping after flattening each image block is used as the initial node feature vector of that node; other implementation methods can also be used in other embodiments, which are not limited here.
[0081] In specific implementation, obtaining the grid coordinates of each image block in the working condition state diagram can be achieved in the following way: after dividing the image blocks, record the row index and column index of each image block in the working condition state diagram; wherein, the row index corresponds to the position on the time axis, from top to bottom as row 1, row 2, etc.; the column index corresponds to the position on the variable axis, from left to right as column 1, column 2, etc.; the above grid coordinates are used to determine the spatial relationship between nodes when constructing the graph structure later; for example, the grid coordinates of the image block located in row i and column j are (i, j), and the row and column indices of each image block are used as the grid coordinates corresponding to that image block; other embodiments can also be implemented in other ways, which are not limited here.
[0082] In specific implementation, taking the node corresponding to each image block as the center, the final connected neighbor nodes of each image block can be determined based on the feature distance between the initial node feature vectors of each node. This can be achieved in the following way: For the node corresponding to the image block, candidate neighbor nodes are screened within a local fan-shaped region centered on the grid coordinates of the node. The local fan-shaped region is limited to the right side of the central node (i.e., the positive time direction), the bottom side (i.e., the positive variable direction), and the lower right diagonal direction, forming a 90° fan. For each direction, nodes 1 to K steps away from the central node are enumerated as candidate neighbors. Then, for all candidate neighbors, the Euclidean distance between their initial node feature vectors and the initial node feature vector of the central node is calculated. Furthermore, all candidate distances are statistically analyzed, and the mean and standard deviation of all candidate distances are calculated. The mean minus the standard deviation is used as a dynamic threshold. If the feature distance of a candidate node is less than the threshold, it is selected as the final neighbor; otherwise, it is excluded. Candidate nodes with feature distances less than the dynamic threshold are taken as the final neighbor nodes of that node. Other implementation methods can also be used in other embodiments, which are not limited here.
[0083] In specific implementation, connecting edges are constructed using all nodes and their corresponding selected neighbor nodes, and all edges are undirected to obtain the operating condition association network of the hydrocracking unit. This can be achieved in the following way: all nodes are treated as graph nodes, and all neighbor relationships selected in step 3 are treated as directed edges, pointing from the central node to the neighbor nodes. To enable the graph structure to support bidirectional information propagation, all directed edges are undirected: if there is an edge from node i to node j, then an edge from node j to node i is added simultaneously. Finally, an undirected graph, i.e., the operating condition association network, is obtained. In this network, nodes have initial node feature vectors, and edges reflect the spatiotemporal dependencies between nodes. For example, a node may be connected to its right neighbor node to capture time evolution; it may be connected to its lower neighbor node to capture variable coupling; or it may be connected to its lower right diagonal node to capture time-varying coupling. All nodes and edges after undirection processing are used as the operating condition association network of the hydrocracking unit. Other embodiments may also use other methods, which are not limited here.
[0084] It should be noted that the working condition association network in this application is a graph structure data constructed by dividing the image blocks in the working condition state graph as nodes and the spatiotemporal dependencies between nodes as edges. It is used to convert image data into graph data for relation modeling, so as to capture the coupling relationship between different sensor variables at different times through graph convolution operations in subsequent steps. At the same time, the fan-shaped neighborhood dynamic graph construction mechanism avoids redundant calculations and noise interference caused by full connection.
[0085] In some embodiments, constructing multi-scale relative difference feature vectors of the working condition association network, and then obtaining enhanced features of multiple nodes under the current working condition mode, can be achieved by the following steps:
[0086] Traverse each edge in the aforementioned working condition association network to obtain the characteristics of the central node and neighboring nodes of each edge;
[0087] Determine the directional difference features and amplitude difference features between the features of the neighboring nodes and the features of the center node.
[0088] By combining interactive coupling features, the corresponding multi-scale relative difference feature vectors are obtained;
[0089] Attention-weighted aggregation and residual connection are performed on all relative difference feature vectors to obtain enhanced node features under the current working condition mode.
[0090] In specific implementation, traversing each edge in the working condition association network to obtain the center node features and neighbor node features of each edge can be achieved in the following way: For each undirected edge in the working condition association network, its two endpoints are taken as the center node and neighbor node, respectively; for each edge, the feature vectors of these two nodes in the current layer are obtained. Initially, these feature vectors are the initial node feature vectors. Then, in the multi-layer propagation of the network, the feature vectors will be updated layer by layer; this step traverses all edges, prepares a pair of node features required for subsequent calculation for each edge, takes the two endpoints of each edge as the center node and neighbor node, respectively, and obtains their corresponding feature vectors; other embodiments can also be implemented in other ways, which are not limited here.
[0091] In specific implementation, determining the directional difference features, amplitude difference features, and interactive coupling features of the neighbor node features and the center node features to obtain the corresponding multi-scale relative difference feature vector can be achieved in the following way: For each edge, let the center node feature be x and the neighbor node feature be y; First, calculate the directional difference feature: y - x, to obtain a vector whose positive and negative signs indicate the direction of change, with positive values indicating that the neighbor is larger than the center and negative values indicating that the neighbor is smaller; then calculate the amplitude difference feature: take the absolute value of each element of y - x to obtain |y - x|, which reflects the magnitude of the change and is independent of the direction; finally, calculate the interactive coupling feature: multiply x and y element by element, i.e., x * y, this feature can characterize the nonlinear interaction between two node features. For example, when both variables increase simultaneously, the product will increase significantly. The above four vectors, namely x itself, yx, |yx|, and x*y, are concatenated in order to form a long vector, which is the multi-scale relative difference feature vector corresponding to the edge. Continue to determine the multi-scale relative difference feature vectors corresponding to the remaining edges. The relative difference feature vector integrates the state of the image patch node itself, the difference direction with its neighbors, the difference magnitude, and the interaction information. The vector obtained by concatenating the direction difference feature, the magnitude difference feature, the interaction coupling feature, and the center node feature is used as the multi-scale relative difference feature vector. Other embodiments can also be implemented in other ways, which are not limited here.
[0092] In specific implementation, attention-weighted aggregation and residual connection are performed on all relative difference feature vectors to obtain enhanced node features under the current working condition mode. This can be achieved in the following way: For each node, collect the multi-scale relative difference feature vectors corresponding to all its neighbor edges; First, calculate the attention score of each neighbor through a lightweight multilayer perceptron. The perceptron takes the relative difference feature vector of the edge and the relative position embedding of the two nodes, such as the grid coordinate difference, as input and outputs a scalar score; Then, perform softmax normalization on the scores of all neighbors to obtain the attention weight of each neighbor. Multiply the multi-scale relative difference feature vector of each neighbor by its attention weight and sum them to obtain the aggregated feature of the central node; Next, map the aggregated feature to the same dimension as the original node feature through a linear transformation layer, then pass it through a non-linear activation function such as GELU, and finally add it to the original central node feature, i.e., perform residual connection to obtain the enhanced node feature; Perform the same operation on all nodes to obtain the enhanced node features of all nodes in the current layer; Other implementation methods can also be used in other embodiments, which are not limited here.
[0093] It should be noted that the enhanced node features in this application include not only the node's own information but also the multi-scale differential information of neighboring nodes, enabling a more sensitive perception of changes in operating conditions.
[0094] In step 104, historical operating data of the hydrocracking unit is acquired, and multiple stable operating condition diagrams for identifying the operating conditions are determined based on the historical operating data.
[0095] In practice, obtaining historical operating data of the hydrocracking unit can be achieved in the following way: First, connect to the core data carriers for unit operation and management, namely industrial-grade data platforms such as Distributed Control System (DCS) and Manufacturing Execution System (MES). Through standardized data interaction interfaces, extract in batches the full-cycle historical operating records covering key process parameters such as reaction temperature, reaction pressure, hydrogen-to-oil ratio, circulating hydrogen purity, feed flow rate and properties, and product yield. For heterogeneous data formats stored in different systems, first complete the unified conversion of data formats. Then, ensure the validity of the data through cleaning methods such as outlier removal, missing value completion, and operational disturbance data filtering. At the same time, according to the preset time granularity, such as minute level, the raw data is normalized to finally obtain the historical operating data of the hydrocracking unit.
[0096] In some embodiments, reference Figure 3 As shown, this figure is an exemplary flowchart for determining a stable operating condition diagram in some embodiments of this application. In this embodiment, determining multiple stable operating condition diagrams based on the historical operating data can be achieved using the following steps:
[0097] First, in step 1041, historical sensor data of the hydrocracking unit under multiple known operating conditions are acquired, and multi-channel image encoding is performed on all historical sensor data to obtain the corresponding historical operating condition state diagram.
[0098] Secondly, in step 1042, block mapping and network construction of the historical working condition state diagram are performed to obtain the historical association network corresponding to the historical sample;
[0099] Finally, in step 1043, multi-scale relative difference feature vector construction and node feature enhancement are performed on all historical correlation networks to obtain multiple stable working condition maps that identify the working condition status.
[0100] In specific implementation, multi-channel image encoding is performed on all historical sensor data to obtain the corresponding historical operating condition state map. This can be achieved in the following way: First, historical multi-source sensor data under multiple known operating conditions during normal operation of the hydrocracking unit are collected; each operating condition mode corresponds to a stable operating state, such as full-load operation, half-load operation, operation under a certain feed ratio, etc.; for each segment of historical data: the data is time-aligned, a numerical matrix, a missing indicator matrix, and a sampling interval matrix are constructed, and these are respectively used as RGB three channels to obtain the historical operating condition state map corresponding to that segment of historical samples. This process is repeated to determine the historical operating condition state maps corresponding to the remaining segments of historical samples. The historical working condition state map performs block mapping and working condition network construction to obtain the historical association network corresponding to the historical sample. This can be achieved in the following way: For each historical working condition state map, the image is first uniformly divided into image blocks, which are then converted into initial node feature vectors through linear mapping, and the grid coordinates of each block are recorded. Then, with each node as the center, neighboring nodes are dynamically selected in a local fan-shaped region based on feature distance. Finally, undirected edges are constructed to obtain the historical association network corresponding to the historical sample. The nodes in the historical association network have initial node features, and the edges reflect the spatiotemporal dependencies in the historical data. Other embodiments may also use other methods to achieve this, which are not limited here.
[0101] In specific implementation, multi-scale relative difference feature vector construction and node feature enhancement are performed on all historical associated networks to obtain multiple stable working condition maps for identifying working conditions. This can be achieved in the following way: For each historical associated network, firstly, the edges in the network are traversed, and multi-scale relative difference feature vectors are calculated. Then, through attention-weighted aggregation and residual connections, the enhanced features of each node are obtained. For multiple historical samples under the same working condition mode, the enhanced feature vectors of all nodes in all samples are aggregated, for example, by calculating the mean of each node position or using attention fusion, to obtain a standard feature template representing the mode. The standard feature templates of all modes are collected, and the standard feature template obtained by statistically aggregating the enhanced node features of all historical samples under the same mode is used as the stable working condition map of the mode. The stable working condition map is used for working condition matching in subsequent real-time monitoring. Other implementation methods can also be used in other embodiments, which are not limited here.
[0102] It should be noted that the historical operating condition map in this application is image data obtained by performing multi-channel image encoding on historical multi-source sensor data of the hydrocracking unit under known stable operating conditions. It is used to construct standard feature templates for each stable operating condition mode, so as to compare the current operating condition features with historical benchmarks in real-time monitoring, thereby accurately determining the current operating status of the unit. At the same time, through statistical condensation of multiple historical samples, it overcomes the random bias that may exist in a single sample, and improves the stability and reliability of operating condition identification.
[0103] In step 105, multiple matching degrees of all stable operating condition maps and all node features are determined. When all matching degrees are lower than a preset threshold, the operating mode of the hydrocracking unit is determined to be a transition state. A blue warning is issued at the beginning stage, and the control parameters are adjusted at the evolution stage to suppress the operating condition fluctuations of the transition state.
[0104] In some embodiments, determining multiple matching degrees for all stable condition atlases and all node features can be achieved using the following steps:
[0105] Obtain the preset steady-state determination threshold;
[0106] Calculate the similarity between the features of each node and all stable operating condition maps, and use all similarities as the corresponding matching degree;
[0107] Compare all matching degrees with the steady-state determination threshold.
[0108] In specific implementation, the similarity between each node feature and all stable operating condition graphs is calculated, and all similarities are used as the corresponding matching degree. This can be achieved in the following way: For all enhanced node feature vectors under the current operating condition mode, the similarity between them and each stable operating condition graph in the stable operating condition graph set is calculated; where the similarity can be measured by cosine similarity or Pearson correlation coefficient, etc.; for each mode, the similarity between all nodes and the mode template (stable operating condition graph) is averaged to obtain a comprehensive similarity value, which is used as the matching degree between the current operating condition and the mode; the same operation is performed on all modes to obtain a set of matching degrees, each matching degree corresponding to a stable operating condition mode, and the comprehensive similarity corresponding to each mode is used as the matching degree of the mode; other methods can also be used in other embodiments, which are not limited here.
[0109] In some embodiments, issuing a blue warning at the initial stage and adjusting control parameters during the evolution stage to suppress transient condition fluctuations can be achieved through the following steps:
[0110] Identify the initial and evolutionary stages of the transition state;
[0111] A blue warning is triggered during the initial phase, and the warning information is pushed to the central control station to remind operators to pay attention to changes in the current operating conditions.
[0112] During the evolution phase, the control parameters of the hydrocracking unit are adjusted based on the changing trends of all relative differential eigenvectors in the operating condition correlation network. These control parameters include reactor inlet temperature, circulating hydrogen flow rate, and high-pressure water injection rate.
[0113] In specific implementation, the identification of the initial and evolutionary stages of the transition state can be achieved in the following way: after determining that the current state is in a transition state, the internal stages of the transition state are identified by using the temporal changes of the constructed multi-scale relative difference feature vector. Specifically, the changing trends of the directional difference feature and the amplitude difference feature among all node features are monitored: when the amplitude difference feature begins to increase steadily for several consecutive moments, but the direction of change is not yet stable, it is marked as the initial stage; when the amplitude difference feature remains large and the direction of change is consistent, such as when the pressure continues to rise, it is marked as the evolutionary stage; when the amplitude difference feature begins to decrease and approaches zero, it is marked as the end point. By continuously monitoring the changing patterns of these features, the various stages of the transition state can be divided in real time. The period when the amplitude difference feature begins to increase and the direction is unstable is taken as the initial stage, and the period when the amplitude difference feature remains large and the direction is consistent is taken as the evolutionary stage. Other embodiments can also be implemented in other ways, which are not limited here.
[0114] In another aspect, in some embodiments, this application provides an intelligent operation monitoring system for petrochemical production plants, with reference to... Figure 4 The figure is a schematic diagram of the structure of an intelligent operation monitoring system for petrochemical production equipment according to some embodiments of this application. The intelligent operation monitoring system for petrochemical production equipment includes: an acquisition module 401, a processing module 402, and an execution module 403, which are described below:
[0115] Acquisition module 401, in this application, is mainly used to collect multi-source sensor data during the operation of hydrocracking unit in petrochemical production;
[0116] Processing module 402, in this application, is used to perform multi-channel image encoding on all numerical variables, data missing masks and sampling time intervals in the multi-source sensor data to construct the operating condition status diagram of the hydrocracking unit during current operation.
[0117] It should be noted that the processing module 402 in this application is also used to perform block mapping on the operating condition diagram to obtain the operating condition association network of the multi-modal operating conditions of the hydrocracking unit, construct the multi-scale relative difference feature vector of the operating condition association network, and then obtain the enhanced multiple node features under the current operating condition mode.
[0118] Additionally, it should be noted that the processing module 402 in this application is also used to acquire historical operating data of the hydrocracking unit and determine multiple stable operating condition diagrams for identifying the operating conditions based on the historical operating data.
[0119] The execution module 403 in this application is mainly used to determine multiple matching degrees of all stable operating condition maps and all node features. When all matching degrees are lower than a preset threshold, the operating mode of the hydrocracking unit is determined to be a transition state, and a blue warning is issued in the initial stage. In the evolution stage, the control parameters are adjusted to suppress the operating condition fluctuations of the transition state.
[0120] In addition, this application also provides a computer device, which includes a memory and a processor. The memory stores code, and the processor is configured to acquire the code and execute the above-described intelligent operation monitoring method for petrochemical production equipment.
[0121] In some embodiments, reference Figure 5 The figure is a schematic diagram of the structure of a computer device for implementing an intelligent operation monitoring method for petrochemical production plants, according to some embodiments of this application. The intelligent operation monitoring method for petrochemical production plants in the above embodiments can... Figure 5 The computer device shown is used to implement this, and the computer device includes at least one processor 501, a communication bus 502, a memory 503, and at least one communication interface 504.
[0122] Processor 501 can be a general-purpose central processing unit (CPU) or an application-specific integrated circuit (ASIC).
[0123] The communication bus 502 can be used to transmit information between the aforementioned components.
[0124] Memory 503 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital versatile optical discs, Blu-ray discs, etc.), magnetic disks or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. Memory 503 may exist independently and be connected to processor 501 via communication bus 502. Memory 503 may also be integrated with processor 501.
[0125] The memory 503 stores program code for executing the scheme of this application, and its execution is controlled by the processor 501. The processor 501 executes the program code stored in the memory 503. The program code may include one or more software modules. The method used in the above embodiments can be implemented by the processor 501 and one or more software modules in the program code in the memory 503.
[0126] Communication interface 504 uses any transceiver-like device to communicate with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc.
[0127] In a specific implementation, as one example, a computer device may include multiple processors, each of which may be a single-core (single-CPU) processor or a multi-core (multi-CPU) processor. Here, a processor may refer to one or more devices, circuits, and / or processing cores used to process data (e.g., computer program instructions).
[0128] The aforementioned computer device can be a general-purpose computer device or a special-purpose computer device. In specific implementations, the computer device can be a desktop computer, a portable computer, a network server, a handheld digital assistant (PDA), a mobile phone, a tablet computer, a wireless terminal device, a communication device, or an embedded device. This application does not limit the type of computer device.
[0129] In addition, this application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described intelligent operation monitoring method for petrochemical production equipment.
[0130] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.
[0131] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.
Claims
1. A method for intelligent operation monitoring of a petrochemical production plant, characterized in that, Includes the following steps: Collect multi-source sensor data during the operation of hydrocracking units in petrochemical production; All numerical variables, data missing masks, and sampling time intervals in the multi-source sensor data are encoded using multi-channel image encoding to construct a current operating condition diagram of the hydrocracking unit. The operating condition diagram is divided into blocks for mapping to obtain the operating condition association network of the multi-modal operating conditions of the hydrocracking unit. The multi-scale relative difference feature vector of the operating condition association network is constructed to obtain the enhanced node features under the current operating condition mode. Acquire historical operating data of the hydrocracking unit, and determine multiple stable operating condition diagrams based on the historical operating data to identify the operating conditions; Multiple matching degrees are determined for all stable operating condition maps and all node features. When all matching degrees are lower than a preset threshold, the operating mode of the hydrocracking unit is determined to be a transition state. A blue warning is issued at the beginning stage, and the control parameters are adjusted at the evolution stage to suppress the operating condition fluctuations of the transition state.
2. The method as described in claim 1, characterized in that, The process of constructing a current operating status diagram of the hydrocracking unit by performing multi-channel image encoding on all numerical variables, data missing masks, and sampling time intervals from the multi-source sensor data specifically includes: Time alignment is performed on the multi-source sensor data; Extract the numerical variables of each sampling time from the aligned multi-source sensor data, obtain the numerical matrix, and input it into the numerical channel; Construct a missing indicator matrix based on the missing distribution of the numerical matrix data and input the missing channels; Determine the sampling interval matrix for the time interval of data acquisition and input the forward hold interval channel; By aligning the numerical channels, missing channels, and interval channels, a current operating status diagram of the hydrocracking unit is obtained.
3. The method as described in claim 1, characterized in that, The operation condition diagram is divided into blocks for mapping to obtain the multimodal operation condition association network of the hydrocracking unit, which specifically includes: The working condition diagram is evenly divided into multiple image blocks, and all image blocks are converted into initial node feature vectors through linear mapping; Obtain the grid coordinates of each image block in the working condition diagram; Using the node corresponding to each image patch as the center, the final connected neighbor nodes of each image patch are determined based on the feature distance between the initial node feature vectors of each node. By constructing connecting edges with all nodes and their corresponding selected neighbor nodes, and then undirecting all edges, the operating condition association network of the hydrocracking unit is obtained.
4. The method as described in claim 1, characterized in that, Constructing the multi-scale relative difference feature vectors of the working condition association network, and then obtaining the enhanced features of multiple nodes under the current working condition mode, specifically includes: Traverse each edge in the aforementioned working condition association network to obtain the characteristics of the central node and neighboring nodes of each edge; The directional difference features, amplitude difference features, and interactive coupling features between the features of the neighboring nodes and the features of the center node are determined to obtain the corresponding multi-scale relative difference feature vectors. Attention-weighted aggregation and residual connection are performed on all relative difference feature vectors to obtain enhanced node features under the current working condition mode.
5. The method as described in claim 1, characterized in that, The multiple stable operating condition diagrams identified based on the historical operating data specifically include: Historical sensor data of the hydrocracking unit under multiple known operating conditions are acquired, and multi-channel image encoding is performed on all historical sensor data to obtain the corresponding historical operating condition state diagram. Perform block mapping and network construction on the historical working condition state diagram to obtain the historical association network corresponding to the historical sample; Multi-scale relative difference feature vector construction and node feature enhancement are performed on all historical correlation networks to obtain multiple stable working condition maps that identify the working condition status.
6. The method as described in claim 1, characterized in that, Determining multiple matching degrees for all stable operating condition atlases and all node features specifically includes: Obtain the preset steady-state determination threshold; Calculate the similarity between the features of each node and all stable operating condition maps, and use all similarities as the corresponding matching degree; Compare all matching degrees with the steady-state determination threshold.
7. The method as described in claim 1, characterized in that, Issuing a blue alert during the initial stage and adjusting control parameters during the evolution stage to suppress transient operating condition fluctuations specifically includes: Identify the initial and evolutionary stages of the transition state; A blue warning is triggered during the initial phase, and the warning information is pushed to the central control station to remind operators to pay attention to changes in the current operating conditions. During the evolution phase, the control parameters of the hydrocracking unit are adjusted based on the changing trends of all relative differential eigenvectors in the operating condition correlation network. These control parameters include reactor inlet temperature, circulating hydrogen flow rate, and high-pressure water injection rate.
8. An intelligent operation monitoring system for petrochemical production facilities, characterized in that, include: The acquisition module is used to collect multi-source sensor data during the operation of hydrocracking units in petrochemical production. The processing module is used to perform multi-channel image encoding on all numerical variables, data missing masks and sampling time intervals in the multi-source sensor data to construct the operating status diagram of the hydrocracking unit during current operation. The processing module is also used to perform block mapping on the operating condition diagram to obtain the operating condition association network of the multi-modal operating conditions of the hydrocracking unit, construct the multi-scale relative difference feature vector of the operating condition association network, and then obtain the enhanced multiple node features under the current operating condition mode. The processing module is also used to acquire historical operating data of the hydrocracking unit and determine multiple stable operating condition diagrams based on the historical operating data to identify the operating conditions. The execution module is used to determine multiple matching degrees of all stable operating condition maps and all node features. When all matching degrees are lower than a preset threshold, the operating mode of the hydrocracking unit is determined to be a transition state. A blue warning is issued at the beginning stage, and the control parameters are adjusted at the evolution stage to suppress the operating condition fluctuations of the transition state.
9. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the intelligent operation monitoring method for petrochemical production units as described in any one of claims 1 to 7.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the intelligent operation monitoring method for petrochemical production plants as described in any one of claims 1 to 7.