Complex industrial process operation state evaluation method based on clustering guide graph Transform model
By adopting a cluster-based guided graph-based Transformer model in complex industrial processes, combining graph neural network and Transformer model to capture the space-time dependence relationship, the problem of insufficient accuracy and stability of operating state evaluation in the existing technology is solved, and a more efficient and reliable evaluation effect is achieved.
Patent Information
- Application Number
- CN202510117665.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-24
- Publication Date
- 2025-05-30
AI Technical Summary
The prior art is difficult to effectively capture the space-time dependence relationships in complex industrial processes, resulting in the limitation of the accuracy and stability of operating state evaluation.
Using the Transformer model based on clustering guided graphs, combined with the graph neural network and the Transformer model, the spatial dependencies are captured through the GCN network, and the Transformer network captures the time dependencies to deeply explore key feature information.
It significantly improves the robustness and reliability of operating status evaluation, and can quickly and accurately evaluate the operating status of industrial processes, and is suitable for high-complexity and large-scale industrial production environments.
Smart Images
Figure CN120067793A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of the operation status evaluation of industrial production processes, and specifically relates to a method for evaluating the operation status of complex industrial processes based on a clustering-guided graph Transformer model. Background Technique
[0002] With the continuous improvement of the requirements for product quality, system performance and economic benefits, modern industrial processes are becoming increasingly complex in terms of structure and automation. The reliability and safety issues of these complex industrial processes have become key considerations in industrial system design. In the initial stage of production, the process operation performance can usually be maintained at an ideal level. However, with the accumulation of process disturbances, parameter drifts and operation errors, the operation performance will gradually decline, resulting in the deviation of the industrial process from the optimal operating point. This deviation may cause unstable fluctuations in product quality, and traditional process monitoring technologies often cannot meet the enterprise's demand for optimal economic benefits. Therefore, timely and accurate operation status evaluation has become an important means to improve product quality and economic benefits. Operation status evaluation is based on continuous monitoring under normal operating conditions to further judge the current operating condition of the system, aiming to optimize the production process and improve the comprehensive benefits. Timely and accurate evaluation of the operation status of industrial processes and guiding production operators to implement effective regulation based on this are of great significance for ensuring the stable and efficient operation of complex industrial processes.
[0003] At present, data-driven methods for evaluating operating states have become the mainstream research direction. For different process characteristics, methods based on multivariate statistics, such as principal component analysis, partial least squares projection to latent structures, etc., have been widely applied. However, in the actual industrial environment, the data often has strong noise interference and exhibits high-dimensional non-linear characteristics. Methods based on multivariate statistics are generally considered shallow learning methods, which are difficult to deeply mine and accurately capture the complex implicit relationships between data. Therefore, there are certain limitations in the accuracy and stability of these methods in operating state evaluation, and it is difficult to meet the urgent needs of modern process industries for efficient, precise management and optimization. With the rapid development of advanced technologies such as sensor technology and the Internet of Things, the amount of data generated in industrial processes has increased significantly. Based on deep learning methods, with their powerful feature extraction and processing capabilities, they have gradually become important tools for solving operating state evaluation problems, such as autoencoders, generative adversarial networks, etc. Although these methods have achieved certain success in specific applications, they are essentially static learning methods, mainly used to solve problems at the data level, and are difficult to cope with the time dependence and dynamic characteristics in complex industrial processes. In a complex industrial process environment, due to continuous and uninterrupted physical and chemical reactions, the data usually has obvious time dependence. Therefore, recurrent neural networks and their variants are widely used to capture the dynamic characteristics and time series features of the data, by deeply mining the deep implicit relationships between data in different states, so as to improve the evaluation accuracy. However, these methods often have difficulty revealing the geometric structures and relationships between multiple variables, and there are certain limitations in using spatial information for modeling. This limits the ability to comprehensively model multivariate time series in process industries based on these methods, affecting the effective representation and evaluation of operating state characteristics. Graph neural networks have shown good performance in some fields because they can capture the spatial dependence relationships between nodes, but when applied to industrial processes, they still face the problem of receptive field limitations. Specifically, GNNs can usually only effectively learn the local dependence relationships between neighboring nodes in the graph, and it is difficult to capture the information of distal nodes. At the same time, GNNs also have certain limitations in using time information. Therefore, the existing methods have shown certain limitations in learning spatio-temporal dependence relationships. For this reason, there is an urgent need to provide an innovative method that combines the advantages of graph neural networks and Transformer models to effectively capture the spatial dependence relationships and time dependence relationships in industrial processes. Summary of the Invention
[0004] Aiming at the problems existing in the above-mentioned prior art, the present invention provides a method for evaluating the operating state of complex industrial processes based on a clustering-guided graph Transformer model. This method can simultaneously capture the spatial dependence relationships and time dependence relationships in industrial processes, and by integrating the advantages of both, deeply mine key feature information, accurately evaluate different operating states, so as to significantly improve the robustness and reliability of operating state evaluation.
[0005] To achieve the above object, the present invention provides a method for evaluating the operating state of a complex industrial process based on a clustering-guided graph Transformer model, including a coal slime flotation industrial system, where the coal slime flotation industrial system includes a main sizing screen, a secondary sizing screen, a conveyor, a mixing tank, a pressure pump, a heavy medium cyclone, a gangue screen, a clean coal screen, a coal slurry tank, a thickener, a coal slurry preparation device, a foam flotation cell, a qualified medium tank, a circulation pump, and a controller;
[0006] The main sizing screen is used to separate raw coal into coarse coal and clean coal by screening; the feed inlet of the secondary sizing screen is connected to the top discharge outlet of the main sizing screen, and is used to remove slime impurities in the clean coal by screening; the feed end of the conveyor is connected to the top discharge outlet of the secondary sizing screen, and is used to output the screened clean coal to the feed inlet of the mixing tank; the feed inlet of the mixing tank is connected to the discharge outlet of the conveyor, and is used to mix the clean coal with the heavy medium suspension to form a mixture; the inlet end of the pressure pump is connected to the discharge outlet of the mixing tank through a pipeline, and is used to output the mixture to the heavy medium cyclone; the feed inlet of the heavy medium cyclone is connected to the outlet end of the pressure pump, and is used to separate low-density materials and high-density materials by using the density difference of materials to obtain overflow and underflow products; the feed inlet of the gangue screen is connected to the high-density discharge outlet of the heavy medium cyclone, and is used to perform dehydration and desliming operations on the high-density materials. Its high-density discharge outlet discharges tailings, and its low-density discharge outlet discharges dilute magnetic medium; the feed inlet of the clean coal screen is connected to the low-density discharge outlet of the heavy medium cyclone, and is used to perform dehydration and desliming operations on the low-density materials. Its high-density discharge outlet discharges clean clean coal, and its low-density discharge outlet discharges dilute magnetic medium; the feed inlet of the coal slurry tank is connected to the high-density discharge outlet of the clean coal screen, and is used to receive clean clean coal and slime. The discharge outlet of the coal slurry tank is connected with a sieve tube. The upper discharge outlet of the sieve outputs coarse-grained clean coal, and its lower discharge outlet outputs slime; the feed inlet of the thickener is connected to the lower discharge outlet of the sieve tube, and is used to increase the concentration of the slime and reduce the moisture content of the slime, so that the pulp density is maintained within the range of stable process operation; the feed inlets of the coal slurry preparation device are respectively connected to the discharge outlet of the thickener, the flotation agent injection pipeline and the water injection pipeline 1, and are used to fully mix the flotation reagent, water and pulp; the feed inlet of the foam flotation cell is connected to the discharge outlet of the coal slurry preparation device, and is used to obtain the final clean coal in the overflow product of the flotation cell by using the different hydrophilicities of different solid particles in the pulp. Its top discharge outlet outputs clean coal, and its bottom discharge outlet outputs tailings; the feed inlets of the qualified medium tank are respectively connected to the low-density discharge outlet of the gangue screen, the low-density discharge outlet of the clean coal screen, the magnetic medium injection pipeline and the water injection pipeline 2, and are used to complete the preparation operation of the heavy medium suspension; the inlet end of the circulation pump is connected to the discharge outlet of the qualified medium tank through a pipeline, and its outlet end is connected to the feed inlet of the mixing tank through a pipeline, and is used to transport the heavy medium suspension to the mixing tank; the controller is respectively connected to the control valves on the main sizing screen, the secondary sizing screen, the conveyor, the mixing tank, the pressure pump, the heavy medium cyclone, the gangue screen, the clean coal screen, the coal slurry tank, the thickener, the coal slurry preparation device, the foam flotation cell, the qualified medium tank, the circulation pump, the flotation agent injection pipeline, the control valve on the water injection pipeline 1, the control valve on the magnetic medium injection pipeline and the control valve on the water injection pipeline 2, and is used to control the actions of each component;
[0007] The operation status evaluation method includes the following steps:
[0008] Step 1: Train a GCN model using an offline dataset and reconstruct the GCN information transfer method using a clustering algorithm to establish a complex industrial process spatial information learning module;
[0009] A1: During the operation of the coal slime flotation industrial system, collect industrial process data using multiple sensors installed at multiple monitoring nodes and store and record them using a database; extract historical industrial process data from the database as an offline dataset X = (x 1 , x 2 ,..., x T ), and at the same time, collect the corresponding comprehensive economic index data Y = (y 1 , y 2 ,..., y T ); where T is the sampling duration, 1 , y 2 ,..., y T ); where m represents the number of variables in each sample; m represents the number of variables in each sample;
[0010] A2: Perform data normalization preprocessing on the X data and Y data according to formula (1), and perform regularization preprocessing on the X data and Y data according to formula (2);
[0011]
[0012]
[0013] In the formula, Z is the input data, min(Z) is the minimum value of the input data, max(Z) is the minimum value of the input data, μ represents the mean of the input data, and σ is the standard deviation of the input data;
[0014] A3: Use a sliding window to divide the offline data (X, Y) to obtain the batch dataset (X N , Y N ) input into the model each time, where N represents the number of samples in a batch, and X N and Y N are calculated through formula (3) and formula (4) respectively;
[0015]
[0016] A4: Use formula (5) to construct its corresponding adjacency matrix A based on the input feature X N to obtain a graph structure representation of the time series data;
[0017]
[0018] In the formula, W1 and W 2 are respectively trainable randomly initialized weights, and d is the dimension of X N to prevent gradient explosion;
[0019] A5: Reconstruct the GCN information passing method using the clustering algorithm k = {1, 2,..., K}, where K is the number of cluster centers; during the process of indexing the cluster center samples, calculate the distance d of the sample from the cluster center according to formula (6) k , and calculate the index of the sample closest to the cluster center according to formula (7)
[0020] d k = ||H k - c k || 2 (6);
[0021]
[0022] In the formula, k represents the category of clustering, H k represents the dimension of the GCN hidden layer, and c k represents the dimension of the cluster center;
[0023] A6: Construct the degree matrix using the cluster center index and dynamically update A to obtain as shown in formula (8) and formula (9);
[0024]
[0025] In the formula, I K is the cluster center connection vector, and λ is a trainable parameter;
[0026] A7: Randomly initialize the GCN parameters, and input the parameters X N and A into the GCN network for spatial feature extraction. The GCN message passing is as shown in formula (6);
[0027]
[0028] In the formula, H (l+1) represents the output of the l + 1 layer in the GCN network, and H l represents the input of the l layer in the GCN network;
[0029] Step 2: Input the node features optimized by GCN and train the Transformer network model to establish an offline model for evaluating the overall operating state;
[0030] B1: Use formula (11) to enhance the node output features of GCN Constructed in an embedded manner H d , and embedded into the Transformer architecture;
[0031]
[0032] where H d is the high-dimensional representation after mapping, and W d represents the linear transformation matrix;
[0033] B2: Embedding position encoding enables the Transformer to capture position information. Among them, the position encoding is obtained according to formulas (12) and (13), and the embedded feature representation H I is obtained;
[0034] PE (pos,2i) = sin(pos / 10000 2i / d )(12);
[0035] PE (pos,2i+1) = cos(pos / 10000 2i / d )(13);
[0036] H I = H d + E pos (14);
[0037] B3: Using the multi-head attention mechanism to capture the long-term dependencies between nodes and fuse the multi-head information output. Among them, the calculation formula of the multi-head attention is shown in formula (15), and the output value H O of the self-attention module is calculated as shown in formula (16);
[0038]
[0039] where Q = H I W Q , K = H I W K and V = H I W V , representing the query matrix, key matrix, and value matrix respectively, and d k represents the dimension of the key vector; Q i , K i and V i represent the query matrix, key matrix, and value matrix of each head; W h represents the linear projection after concatenating the multi-heads;
[0040] B4: Using residual connection and normalization to provide the model with linear transformation ability, as shown in formulas (17) and (18);
[0041]
[0042] B5: Use a multi-layer perceptron to evaluate the operating state of complex industrial processes, as shown in formula (19);
[0043]
[0044] In the formula, W 1 and W 2 represent trainable mapping weights, b 1 and b 2 represent bias vectors, is the final output;
[0045] B6: Use the normalized exponential function to normalize the posterior probability, as shown in formula (20);
[0046]
[0047] B7: Calculate the classification loss function L(p, y) according to formula (21);
[0048]
[0049] B8: Reverse fine-tune the model parameters until the classification loss error is minimized, and train to obtain the complete operating state evaluation model CGGformer;
[0050] Step 3: Use online data for operating state evaluation;
[0051] C1: Use the monitoring sensor group installed in the coal slime flotation industrial system to perform online sampling to obtain online process data X, and perform normalization and standardization preprocessing on X;
[0052] C2: Sample the online data using a sliding window to obtain sequence data of the same length;
[0053] C3: Construct a graph structure representation of the sequence data and input it into the trained GCN model to obtain input features;
[0054] C4: Input the node features optimized by GCN into the trained Transformer model to obtain the posterior probability p of the online data x t at different times t;
[0055] C5: Determine the operating state with the highest probability according to the posterior probability, which is the final online operating state evaluation result of x t ;
[0056] The present invention proposes an innovative method combining the advantages of graph neural network and Transformer model, which can effectively learn the key features of hidden state and dynamic correlation between comprehensive economic indicators by using GCN network and Tranformer network to capture the spatial dependency and time dependency of data and process in complex industrial process respectively. For GCN network, clustering algorithm is used to reconstruct the information transmission expression mode, and a global spatial information expression mode is constructed. The optimized node feature encoding is input into Transformer network, and the time dependency between nodes is captured by self-attention mechanism. During the training process, the output result of the model is combined with comprehensive economic indicators for supervised learning. By classifying the performance level of process offline data according to the comprehensive economic indicators of the industrial process, the model can learn the key feature expressions related to these economic indicators. The final loss is calculated according to the comprehensive economic indicators of the industrial process and the model input, and the model is reversely fine-tuned using the loss function to optimize the model parameters until the classification loss error is minimized, thereby obtaining a high-precision process operation status evaluation model. The process operation status evaluation model can represent the industrial process data structure information in a graph structure, effectively extract the spatial dependency and time dependency of complex industrial process data and process, and improve the accuracy and robustness of process operation status evaluation. Through real-time monitoring and accurate evaluation of process status, it can provide important decision-making support for production operators and optimize the control strategy of the production process, thereby improving the overall operating efficiency of the industrial process and reducing energy and resource waste.
[0057] This method can simultaneously capture the spatial dependency and temporal dependency in the industrial process, and by integrating the advantages of both, it can deeply mine key feature information and accurately evaluate different operating states, thereby significantly improving the robustness and reliability of operating state evaluation. The advantage of this method is that it can quickly and accurately evaluate the operating state of industrial processes, and is particularly suitable for highly complex and large-scale industrial production environments. At the same time, this method is conducive to the accurate and stable evaluation of the operating state of complex industrial processes, ensuring the safe, stable and efficient operation of industrial processes, and can effectively guarantee the quality of industrial products, ensure the consistency and stability of products, provide a strong guarantee for enterprises to improve their comprehensive economic benefits, and help promote enterprises to develop in the direction of efficiency and intelligence. BRIEF DESCRIPTION OF THE DRAWINGS
[0058] Figure 1 It is the overall flow chart of the industrial process operation status evaluation method in the present invention;
[0059] Figure 2 It is a schematic diagram of the CGGformer structure in the present invention;
[0060] Figure 3It is a bar chart of the accuracy rate of the operation status evaluation of complex industrial processes based on the CGGformer model in the present invention;
[0061] Figure 4 It is a schematic diagram of the online evaluation result of the operation status evaluation of complex industrial processes based on the CGGformer model in the present invention;
[0062] Figure 5 It is a schematic diagram of the feature visualization result of the operation status evaluation of complex industrial processes based on the CGGformer model in the present invention;
[0063] Figure 6 It is a confusion matrix of the operation status evaluation of complex industrial processes based on the CGGformer model in the present invention;
[0064] Figure 7 It is a schematic diagram of the structure of the coal slime flotation industrial system in the present invention. Detailed implementation manners
[0065] The present invention provides a method for evaluating the operation status of complex industrial processes based on a clustering-guided graph (CGGformer, Cluster-guided graph Transformer)
[0066] Transformer model, including a coal slime flotation industrial system, as Figure 7 shown, the coal slime flotation industrial system includes a main sizing screen, a secondary sizing screen, a conveyor, a mixing tank, a pressure pump, a heavy medium cyclone, a gangue screen, a clean coal screen, a coal slurry tank, a thickener, a coal slurry preparation device, a foam flotation cell, a qualified medium tank, a circulating pump, and a controller;
[0067] The main sizing screen is used to separate raw coal into coarse coal and clean coal by screening; the feed inlet of the secondary sizing screen is connected to the top discharge outlet of the main sizing screen, and is used to remove slime impurities in the clean coal by screening; the feed end of the conveyor is connected to the top discharge outlet of the secondary sizing screen, and is used to output the screened clean coal to the feed inlet of the mixing tank; the feed inlet of the mixing tank is connected to the discharge outlet of the conveyor, and is used to mix the clean coal with the heavy medium suspension to form a mixture; the inlet end of the pressure pump is connected to the discharge outlet of the mixing tank through a pipeline, and is used to output the mixture to the heavy medium cyclone; the feed inlet of the heavy medium cyclone is connected to the outlet end of the pressure pump, and is used to separate low-density materials and high-density materials by using the density difference of materials to obtain overflow and underflow products; the feed inlet of the gangue screen is connected to the high-density discharge outlet of the heavy medium cyclone, and is used to perform dehydration and desliming operations on the high-density materials. Its high-density discharge outlet discharges tailings, and its low-density discharge outlet discharges dilute magnetic medium; the feed inlet of the clean coal screen is connected to the low-density discharge outlet of the heavy medium cyclone, and is used to perform dehydration and desliming operations on the low-density materials. Its high-density discharge outlet discharges clean clean coal, and its low-density discharge outlet discharges dilute magnetic medium; the feed inlet of the coal slurry tank is connected to the high-density discharge outlet of the clean coal screen, and is used to receive clean clean coal and slime. The discharge outlet of the coal slurry tank is connected with a sieve tube. The upper discharge outlet of the sieve outputs coarse-grained clean coal, and its lower discharge outlet outputs slime; the feed inlet of the thickener is connected to the lower discharge outlet of the sieve tube, and is used to increase the concentration of the slime and reduce the moisture content of the slime, so that the pulp density is maintained within the range of stable process operation; the feed inlet of the coal slurry preparation device is respectively connected to the discharge outlet of the thickener, the flotation agent injection pipeline and the water injection pipeline 1, and is used to fully mix the flotation agent, water and pulp; the feed inlet of the foam flotation cell is connected to the discharge outlet of the coal slurry preparation device, and is used to obtain the final clean coal in the overflow product of the flotation cell by using the different hydrophilicities of different solid particles in the pulp. Its top discharge outlet outputs clean coal, and its bottom discharge outlet outputs tailings; the feed inlet of the qualified medium tank is respectively connected to the low-density discharge outlet of the gangue screen, the low-density discharge outlet of the clean coal screen, the magnetic medium injection pipeline and the water injection pipeline 2, and is used to complete the preparation operation of the heavy medium suspension; the inlet end of the circulation pump is connected to the discharge outlet of the qualified medium tank through a pipeline, and its outlet end is connected to the feed inlet of the mixing tank through a pipeline, and is used to transport the heavy medium suspension to the mixing tank; the controller is respectively connected to the control valves on the main sizing screen, secondary sizing screen, conveyor, mixing tank, pressure pump, heavy medium cyclone, gangue screen, clean coal screen, coal slurry tank, thickener, coal slurry preparation device, foam flotation cell, qualified medium tank, circulation pump, flotation agent injection pipeline, control valve on the water injection pipeline 1, control valve on the magnetic medium injection pipeline, control valve on the water injection pipeline 2, and is used to control the actions of each component;
[0068] The present invention provides a method for evaluating the operating status of a complex industrial process based on a clustering guided graph Transformer model. The CGGformer (based on a clustering guided graph Transformer model) is a method combining a graph neural network and a Transformer network, which comprises the following steps:
[0069] Step 1: Use offline data sets to train the GCN model and use clustering algorithms to reconstruct the GCN information transmission method and establish a complex industrial process spatial information learning module;
[0070] A1: During the operation of the coal slime flotation industrial system, multiple sensors installed at multiple monitoring nodes are used to collect industrial process data, and the data is stored and recorded in a database; historical industrial process data is extracted from the database as an offline data set X = (x 1 ,x 2 ,...,x T ), at the same time, collect and offline data X = (x 1 ,x 2 ,...,x T ) corresponding to the comprehensive economic indicator data Y=(y 1 ,y 2 ,...,y T ), where T is the sampling time, m represents the number of variables in each sample;
[0071] A2: Perform data normalization preprocessing on X data and Y data according to formula (1), and perform regularization preprocessing on X data and Y data according to formula (2);
[0072]
[0073]
[0074] In the formula, Z is the input data, min(Z) is the minimum value of the input data, max(Z) is the minimum value of the input data, μ represents the mean of the input data, and σ is the standard deviation of the input data;
[0075] A3: Use a sliding window to divide the offline data (X, Y) to obtain the batch data set (X N ,Y N ), where N represents the number of samples in a batch, X N and Y N Calculate by formula (3) and formula (4) respectively;
[0076]
[0077] A4: Construct its corresponding adjacency matrix A according to the input feature X using formula (5) to obtain the graph structure representation of the time series data; N
[0078]
[0079] In the formula, W 1 and W 2 are respectively randomly initialized trainable weights, and d is the dimension of X N to prevent gradient explosion;
[0080] A5: Reconstruct the GCN information transmission method using the clustering algorithm k = {1, 2,..., K}, where K is the number of cluster centers; during the process of indexing the cluster center samples, calculate the distance d of the sample from the cluster center according to formula (6) k , and calculate the sample index closest to the cluster center according to formula (7)
[0081] d k = ||H k - c k || 2 (6);
[0082]
[0083] In the formula, k represents the category of clustering, H k represents the dimension of the GCN hidden layer, and c k represents the dimension of the cluster center;
[0084] A6: Construct the degree matrix using the cluster center index and dynamically update A to obtain as shown in formula (8) and formula (9);
[0085]
[0086] In the formula, I K is the cluster center connection vector, and λ is a trainable parameter;
[0087] A7: Randomly initialize the GCN parameters, and input the parameters X N and A into the GCN network for spatial feature extraction. The GCN message passing is as shown in formula (6);
[0088]
[0089] In the formula, H (l+1) represents the output of the (l + 1)-th layer in the GCN network, and H l represents the input of the l-th layer in the GCN network;
[0090] Step 2: Input the node features optimized by GCN into the Transformer network model for training, and establish an offline model for evaluating the overall operating state;
[0091] B1: Use formula (11) to construct the enhanced node output features of GCN into the embedding form H d , and embed it into the Transfomer architecture;
[0092]
[0093] where H d is the high-dimensional representation after mapping, and W d represents the linear transformation matrix;
[0094] B2: Embed the position encoding so that Transformer can capture the position information. Among them, the position encoding is obtained according to formula (12) and formula (13), and the embedded feature representation H I is obtained according to formula (14);
[0095] PE (pos,2i) = sin(pos / 10000 2i / d )(12);
[0096] PE (pos,2i+1) = cos(pos / 10000 2i / d )(13);
[0097] H I = H d + E pos (14);
[0098] B3: Use the multi-head attention mechanism to capture the long-term dependence relationship between nodes and fuse the multi-head information output. Among them, the calculation formula of multi-head attention is shown in formula (15), and the output value H O of the self-attention module is calculated as shown in formula (16);
[0099]
[0100] where Q = H I W Q , K = H I W K and V = H I W V , representing the query matrix, key matrix and value matrix respectively, and d k represents the dimension of the key vector; Q i , K i and V iDenote the query matrix, key matrix, and value matrix for each head; W h Denote the linear projection after concatenating multiple heads;
[0101] B4: Use residual connections and normalization to provide the model with the ability of linear variation, as shown in Formulas (17) and (18);
[0102]
[0103] B5: Use a multi-layer perceptron for the operation state evaluation of complex industrial processes, as shown in Formula (19);
[0104]
[0105] In the formula, W 1 and W 2 Denote the trainable mapping weights, b 1 and b 2 Denote the bias vectors, is the final output;
[0106] B6: Use the softmax function to normalize the posterior probability, as shown in Formula (20);
[0107]
[0108] B7: Calculate the classification loss function L(p, y) according to Formula (21);
[0109]
[0110] B8: Backward fine-tune the model parameters until the classification loss error is minimized, and train to obtain the complete operation state evaluation model CGGformer;
[0111] Step 3: Use online data for operation state evaluation;
[0112] C1: Use the monitoring sensor group installed in the coal slime flotation industrial system to perform online sampling to obtain online process data X, and perform normalization and standardization preprocessing on X;
[0113] C2: Sample the online data using a sliding window to obtain sequence data with consistent lengths;
[0114] C3: Construct a graph structure representation of the sequence data and input it into the trained GCN model to obtain input features;
[0115] C4: Input the node features optimized by GCN into the trained Transformer model to obtain the posterior probability p of the online data x t at different times t;
[0116] C5: Determine the operating state with the highest probability based on the posterior probability, which is x t The final online operating state evaluation result.
[0117] The technical solution of the present invention will be described in detail below in conjunction with embodiments, and its feasibility will be verified.
[0118] Embodiment:
[0119] The coal slime flotation process is a typical process industry process operating in a harsh open environment. There are various interference and uncertainty factors in the operating environment, which often make the current process monitoring and operating state evaluation methods etc. not comprehensively perceive the working condition information. Aiming at the problem that it is difficult to capture the spatio-temporal information of the data in the coal slime flotation process, the present invention uses CGGformer to capture the spatio-temporal dependence relationship related to the data and the process, can more effectively extract the implicit features of different operating states, and evaluate the current operating state more accurately and robustly.
[0120] The coal slime flotation system in the present invention is as Figure 7 shown, including main links such as raw coal screening, coal medium mixing, primary separation by hydrocyclone, coal slime desliming, coal slime thickening, flotation pretreatment, and foam flotation. Specifically, after the raw coal is crushed and screened, it is mixed with the heavy medium liquid in the mixing tank to form a coal medium mixture, and then sent to the heavy medium hydrocyclone for primary separation. Subsequently, after the separated product is deslimed and dewatered, the coarse clean coal discharged from the clean coal magnetic separator is mixed with the secondary coal slime from other circuits, and sent to the vibrating arc screen for diversion according to the particle size. The coal slime liquid of the fine particle clean coal under the screen is thickened in the thickener, and the coal slime liquid discharged from the underflow port is sent to the pulp preprocessor for "pre-mineralization". Finally, it enters the flotation cell for foam flotation. The overflow from the flotation cell is scraped out by a mechanical scraper and becomes clean coal after dehydration and drying. The gangue that cannot float stays at the bottom of the flotation cell as tailings with the pulp and is discharged.
[0121] The ash content of the flotation cell overflow is usually used as a basic index for monitoring the coal quality. The higher the ash content index, the worse the coal quality. In the coal slime flotation process, due to the involvement of a large number of equipment and instruments, the coal slime flotation process has strong non-linear characteristics. To ensure the coal quality, the coal preparation process parameters must be continuously detected and adjusted in actual operation. According to the process mechanism and actual situation, 24 process variables are determined, and the ash content index is determined as the output variable. The specific process variables are shown in Table 1.
[0122] Table 1: Selection of process operating variables
[0123]
[0124] 1) Establish an offline state evaluation model
[0125] Collect the off-line data of the coal slime flotation process. After removing outliers and aligning the data, the remaining samples are 20,000 groups. Select 14,000 groups of data as the off-line data of the coal slime flotation process, and 14,001 - 20,000 as the on-line test data.
[0126] Combined with expert experience, divide the state level of the coal preparation process according to the ash content in the clean coal product. The division of the operating state level and label settings are shown in Table 2.
[0127] Table 2: Division of operating state level and label settings
[0128]
[0129] Set the length of the time series to 100 and set the size of the input batch to 20. During the process of training the model, the cross-entropy loss function is used as the loss function, and the Adam optimization algorithm is used to minimize this loss.
[0130] 2) On-line test of the industrial process operating state evaluation method based on CGGformer
[0131] The structure of the CGGformer model is as Figure 2 shown, which includes two parts: the GCN network and the Transformer network. Figure 3 For the evaluation accuracy results of the operating state of the complex industrial process based on the CGGformer model, the weighted average of the precision is 93.75%, the weighted average of the recall is 93.60%, the weighted average of the F1-score is 93.67% and the weighted average of the accuracy is 93.54%. Combined with Figure 4 It can be seen from the on-line evaluation results of the industrial process operating state evaluation based on the CGGfomer model that the grade results given have a high degree of coincidence with the true state grade of the current operating state, that is, the operating state evaluation method based on CGGformer can effectively identify the state grade to which the operating data of the coal slime flotation process belongs.
[0132] Figure 5 Shows the visualization results of the operating state evaluation features of the complex industrial process based on the CGGformer model. Each category has high distinctiveness, and the decision boundaries between categories are clear, indicating that the classification effect of the model is excellent. Thus, it can be seen that the present invention has high robustness for identifying each state level.
[0133] It can be seen from the above simulation examples that the present invention has a certain degree of effectiveness. By implementing timely and robust operating state evaluation for the coal slime flotation process, it can provide an important guiding basis for the production adjustment of the coal slime flotation process.
[0134] The present invention proposes an innovative method that combines the advantages of graph neural networks and Transformer models. By using GCN networks and Transformer networks to capture the spatial and temporal dependencies of data and processes in complex industrial processes respectively, it can effectively learn the key features that are dynamically related to the hidden state and comprehensive economic indicators. For the GCN network, a clustering algorithm is used to reconstruct the information transfer expression and construct a global spatial information expression. The optimized node feature encoding is input into the Transformer network, and the self-attention mechanism is used to capture the temporal dependencies between nodes. During the training process, the output results of the model are combined with the comprehensive economic indicators for supervised learning. By classifying the performance levels of the process offline data according to the comprehensive economic indicators of the industrial process, the model can learn the key feature expressions related to these economic indicators. The final loss is calculated based on the comprehensive economic indicators of the industrial process and the model input, and the loss function is used to perform backpropagation fine-tuning on the model to optimize the model parameters until the classification loss error is minimized, thereby obtaining a high-precision process operating state evaluation model. This process operating state evaluation model can represent the industrial process data structure information in a graph structure, effectively extract the spatial and temporal dependencies of complex industrial process data and processes, and improve the accuracy and robustness of the process operating state evaluation. Through real-time monitoring and accurate assessment of the process state, it can provide important decision-making support for production operators, optimize the control strategy of the production process, thereby improving the overall operating efficiency of the industrial process and reducing energy and resource waste.
[0135] This method can capture both the spatial and temporal dependencies in industrial processes and, by integrating the advantages of both, deeply excavate key feature information and accurately evaluate different operating states, thus significantly improving the robustness and reliability of the operating state evaluation. The advantage of this method is that it can quickly and accurately evaluate the operating state of industrial processes, especially suitable for high-complexity and large-scale industrial production environments. At the same time, this method is conducive to achieving accurate and stable evaluation of the operating state of complex industrial processes, ensuring the safe, stable and efficient operation of industrial processes, effectively guaranteeing the quality of industrial products, ensuring the consistency and stability of products, providing a strong guarantee for enterprises to improve comprehensive economic benefits, and helping to promote the development of enterprises towards high efficiency and intelligence.
Claims
1. A complex industrial process operation status evaluation method based on a clustering guided graph Transformer model, including a coal slime flotation industrial system, wherein the coal slime flotation industrial system includes a primary grading screen, a secondary grading screen, a conveyor, a mixing barrel, a pressure pump, a heavy medium cyclone, a gangue screen, a clean coal screen, a coal slurry barrel, a thickener, a coal slurry preparation device, a froth flotation tank, a qualified medium barrel, a circulation pump and a controller; The main grading screen is used to separate the raw coal into coarse coal and clean coal by screening; the feed port of the secondary grading screen is connected to the top discharge port of the main grading screen, and is used to remove coal slime impurities in the clean coal by screening; the feed end of the conveyor is connected to the top discharge port of the secondary grading screen, and is used to output the screened clean coal to the feed port of the mixing barrel; the feed port of the mixing barrel is connected to the discharge port of the conveyor, and is used to mix the clean coal with the heavy medium suspension to form a mixture; the inlet end of the pressure pump is connected to the discharge port of the mixing barrel through a pipeline, and is used to output the mixture to the heavy medium cyclone; the feed port of the heavy medium cyclone is connected to the outlet end of the pressure pump, and is used to separate the low-density material and The high-density material is separated to obtain overflow and underflow products; the feed port of the gangue screen is connected to the high-density discharge port of the heavy medium cyclone, which is used to dehydrate and remove the medium of the high-density material, and its high-density discharge port discharges tailings, and its low-density discharge port discharges dilute magnetic media; the feed port of the clean coal screen is connected to the low-density discharge port of the heavy medium cyclone, which is used to dehydrate and remove the medium of the low-density material, and its high-density discharge port discharges clean clean coal, and its low-density discharge port discharges dilute magnetic media; the feed port of the coal slurry barrel is connected to the high-density discharge port of the clean coal screen, which is used to receive clean clean coal and coal slime, and the discharge port of the coal slurry barrel is connected with a screen tube, the upper discharge port of the screen outputs coarse-particle clean coal, and the lower discharge port outputs coal slime; The feed port of the compressor is connected to the lower discharge port of the screen tube to increase the concentration of the coal slime and reduce the moisture content of the coal slime, so that the pulp density is maintained within the range of stable process operation; the feed port of the coal slurry preparation device is respectively connected to the discharge port of the concentrator, the flotation agent filling pipeline and the water filling pipeline, so as to fully mix the flotation agent, water and the pulp; the feed port of the froth flotation tank is connected to the discharge port of the coal slurry preparation device, so as to obtain the final clean coal in the overflow product of the flotation tank by utilizing the different hydrophilicity of different solid particles in the pulp, and the top discharge port outputs clean coal, and the bottom discharge port outputs tailings; the feed port of the qualified medium barrel is respectively connected to the low-density discharge port of the gangue screen and the low-density discharge port of the clean coal screen. The first port, the magnetic medium filling pipeline, and the second water filling pipeline are connected to complete the preparation of the heavy medium suspension; the inlet end of the circulating pump is connected to the discharge port of the qualified medium barrel through a pipeline, and the outlet end is connected to the feed port of the mixing barrel through a pipeline, so as to transport the heavy medium suspension to the mixing barrel; the controller is respectively connected to the main grading screen, the secondary grading screen, the conveyor, the mixing barrel, the pressure pump, the heavy medium cyclone, the gangue screen, the clean coal screen, the coal slurry barrel, the concentrator, the coal slurry preparation device, the foam flotation tank, the qualified medium barrel, the circulating pump, the control valve on the flotation agent filling pipeline, the control valve on the water filling pipeline one, the control valve on the magnetic medium filling pipeline, and the control valve on the water filling pipeline two, so as to control the actions of each component; It is characterized in that The following steps are involved: Step 1: Use offline data sets to train the GCN model and use clustering algorithms to reconstruct the GCN information transmission method and establish a complex industrial process spatial information learning module; A1: During the operation of the coal slime flotation industrial system, multiple sensors installed at multiple monitoring nodes are used to collect industrial process data, and the data is stored and recorded in a database. The historical industrial process data is extracted from the database as an offline data set X = (x1, x2, ..., x T ), at the same time, collect and offline data X=(x1,x2,...,x T ) corresponding to the comprehensive economic indicator data Y=(y1,y2,...,y T ), where T is the sampling time, m represents the number of variables in each sample; A2: Perform data normalization preprocessing on X data and Y data according to formula (1), and perform regularization preprocessing on X data and Y data according to formula (2); In the formula, Z is the input data, min(Z) is the minimum value of the input data, max(Z) is the minimum value of the input data, μ represents the mean of the input data, and σ is the standard deviation of the input data; A3: Use a sliding window to divide the offline data (X, Y) to obtain the batch data set (X N ,Y N ), where N represents the number of samples in a batch, X N and Y N Calculate by formula (3) and formula (4) respectively; A4: Use formula (5) according to the input feature X N Construct the corresponding adjacency matrix A to obtain the graph structure representation of the time series data; Where W1 and W2 are trainable random initialization weights, and d is X N The dimension of is used to prevent gradient explosion; A5: Use clustering algorithm to reconstruct the GCN information transmission mode k = {1, 2, ..., K}, where K is the number of cluster centers; in the cluster center sample indexing process, the distance d between the sample and the cluster center is calculated according to formula (6) k , according to formula (7), calculate the sample index closest to the cluster center d k =‖H k -c k ‖2(6); In the formula, k represents the clustering category, H k represents the hidden layer dimension of GCN, c k Represents the cluster center dimension; A6: Use the cluster center index to construct the degree matrix and dynamically update A to obtain As shown in formula (8) and formula (9); In the formula, I K is the cluster center connection vector, and λ is a trainable parameter; A7: Randomly initialize GCN parameters and set the parameters X N and A are input into the GCN network for spatial feature extraction, and the GCN message transmission is shown in formula (6); In the formula, H (l+1) represents the output of layer l+1 in the GCN network, H l Represents the input of layer l in the GCN network; Step 2: Input the node features optimized by GCN and train the Transformer network model to establish an offline model for overall operation status evaluation; B1: Use formula (11) to enhance the node output features of GCN Constructed into embedded mode H d , and embedded into the Transfomer architecture; In the formula, H d for The high-dimensional representation after mapping, W d represents a linear transformation matrix; B2: Embed position coding to enable Transformer to capture position information. The position coding is obtained according to formula (12) and formula (13), and the embedded feature representation H is obtained according to formula (14). I ; ON (pos,2i) =sin(pos / 10000 2i / d )(12); ON (pos,2i+1) =cos(pos / 10000 2i / d )(13); H I =H d +E pos (14); B3: Use the multi-head attention mechanism to capture the long-term dependency between nodes and fuse the multi-head information output. The calculation formula of the multi-head attention is shown in formula (15). The output value of the self-attention module H O The calculation formula is shown in formula (16); Where Q = H I W Q , K=H I W K and V = H I W V , respectively represent the query matrix, key matrix and value matrix, d k represents the dimension of the key vector; Q i , K i and V i represents the query matrix, key matrix and value matrix of each head; W h It represents the linear projection after splicing multiple heads; B4: Use residual connection and normalization to provide the model with linear change capability, as shown in formula (17) and formula (18); B5: Use multi-layer perceptron to evaluate the operation status of complex industrial processes, as shown in formula (19); Where W1 and W2 represent trainable mapping weights, b1 and b2 represent bias vectors, is the final output; B6: Use the normalized exponential function to normalize the posterior probability, as shown in formula (20); B7: Calculate the classification loss function L(p,y) according to formula (21); B8: Reversely fine-tune the model parameters until the classification loss error is minimized, and train the complete running status evaluation model CGGformer; Step 3: Use online data to evaluate the operation status; C1: Use the monitoring sensor group installed in the coal slime flotation industrial system to perform online sampling to obtain online process data X, and perform normalization and standardization preprocessing on X; C2: Use a sliding window to sample online data to obtain sequence data with consistent length; C3: Construct a graph structure representation of the sequence data and input it into the trained GCN model to obtain input features; C4: Input the node features optimized by GCN into the trained Transformer model to obtain online data x at different times t t The posterior probability p of C5: Determine the operating state with the highest probability based on the posterior probability, which is x t The final online operation status evaluation results.