A cement production tail gas monitoring method and system based on deep learning

Abstract

The application belongs to the field of cement production tail gas monitoring, and discloses a cement production tail gas monitoring method and system based on deep learning, which comprises the following steps: obtaining process detection data sets of cement production; pre-processing and regularizing the data sets; outputting predicted F-Cao content data by using a TA-GCN model based on a time attention graph convolution network (TA-GCN); and judging the equipment operating condition based on the F-Cao content data. The application uses deep learning technology and the process data actually collected in the cement production process to train a prediction model, realizes accurate prediction of the free calcium oxide content, judges whether the equipment has a fault and whether the production process is stable and efficient according to auxiliary variables and prediction variables, and once there is a fault omen in the production process, timely reminds the operator to reset the control parameters, avoids kiln shutdown or long-time non-stable operation due to the fault, and reduces the additional waste gas emission caused by incomplete combustion of raw materials.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of cement production tail gas monitoring, and particularly relates to a cement production tail gas monitoring method and system based on deep learning. BACKGROUND

[0002] At present, with the increasingly prominent global warming problem, carbon emission problem has become the focus of attention in various industries. As the second largest carbon emission industry in the industrial field, the cement industry, the application designs a cement production tail gas monitoring system based on deep learning through experiments, trains a prediction model by using deep learning technology and actual process data collected in the cement production process, aims to realize accurate prediction of the difficult-to-measure variable, i.e., free calcium oxide content, by measuring the easily-measured auxiliary variables such as feeding amount and equipment temperature in the cement production process. And according to the auxiliary variables and the predicted variables, it is judged whether the equipment fails and the production process is stable and efficient. Once there is a sign of failure in the production process, the operator is reminded to reset the control parameters in time, so as to avoid kiln shutdown or long-term non-stable operation due to failure. In this way, the additional exhaust emission caused by incomplete combustion of raw materials is reduced.

[0003] In recent years, graph neural network (GNN) as a branch of deep learning has received extensive attention and has shown excellent learning ability in multiple fields. GNN can propagate information in a graph through a pre-defined structure, which helps to understand the relationship between different points. GCN has strong representation ability. Time series modeling based on GCN can be regarded as spatiotemporal data mining.

[0004] Signed directed graph (SDG) is a graph theory method for qualitatively representing causal relationships. SDG expresses the qualitative influence relationship between system variables in a causal manner as a graph model of directed branches between nodes. The graph describes the physical connectivity between units and loops. From the perspective of information transmission, different nodes generate or carry information about the process. The directed edge describes the information transmission or the direction of information transmission. Compared with complex first-principle mathematical models, the graph only focuses on the connection between different parts of the process and can qualitatively represent the relationship between different variables.

[0005] The application innovatively proposes a cement production process monitoring method based on deep learning, aiming to reduce the additional tail gas generated due to incomplete combustion in the production process. First, the proposed work performs soft measurement on the input process data to obtain the F-CaO content. Then, the F-CaO content and the process data are combined to determine the running state of the production process. Once it is determined to be unstable or even failure, the fault variable is located to assist the staff in troubleshooting and reduce the time of incomplete combustion, thereby achieving the purpose of energy saving and emission reduction.

[0006] Through the above analysis, the problems and defects of the prior art are:

[0007] (1) The detection of F-CaO content in traditional cement production tail gas lacks timeliness, which is not conducive to the safety of cement production.

[0008] (2) The selection of auxiliary variables is limited by the on-site conditions, and manual measurement also has certain error influence. Although the current method can extract different types of data features, it cannot accurately predict the F-CaO content. SUMMARY

[0009] In view of the problems existing in the prior art, the present application provides a cement production tail gas monitoring method and system based on deep learning.

[0010] The present application is implemented in the following way: a cement production tail gas monitoring method based on deep learning, which comprises:

[0011] Step 1: Obtain the process detection data set of cement production.

[0012] Step 2: Preprocess the data set.

[0013] Step 3: Output the predicted F-Cao content data based on the time attention graph convolution network TA-GCN model.

[0014] Step 4: Determine the equipment operating condition based on the F-Cao content data.

[0015] Further, the detection data set is collected by various sensors in the production equipment, and the auxiliary variables are used as inputs of the model.

[0016] Further, the preprocessing of step 2 is a regularization processing, and the processing process is as follows:

[0017] Standardize the data using the MinMaxScaler method. The normalized data X scaled is:

[0018]

[0019] X scaled = X std *(max-min)+min

[0020] Where X is the data to be normalized, X min is a row vector composed of the minimum value in each column of the data X, and X maxA row vector composed of the maximum value in each column of the data X, max is the maximum value of the interval to be mapped, min is the minimum value of the interval to be mapped, the interval is [0, 1].

[0021] Further, the TA-GCN model based on time attention graph convolution network includes feature selection based on maximum information coefficient MIC, a plurality of time attention graph convolution layers TA-GCL and an FC full connection layer.

[0022] The process of the model is:

[0023] The input data is first screened, then the data is aggregated in a graph structure, a plurality of TA-GCLs are stacked to establish a hierarchical model to capture complex spatial-temporal correlation, and finally the data is output through the FC full connection layer.

[0024] Further, the maximum information coefficient MIC is the result of normalizing the maximum value of mutual information MI, the MI is obtained from the joint probability distribution p(J, U) of two continuous random variables J and U, and the expression of the MI is:

[0025]

[0026] The expression of the MIC is:

[0027]

[0028] Wherein, B = T 0.55 , T represents the time step of the data.

[0029] Further, the TA-GCL includes GCL and multi-head attention mechanism MHA based on h-Head;

[0030] First, the data is aggregated in the spatial dimension according to the topological graph, and the multi-head attention mechanism MHA based on h-Head is used to establish long-term dependence relationship in the time dimension;

[0031] Secondly, the residual connection is introduced, the input and output dimensions of each TA-GCL are uniformly set, and the number of heads of each TA-GCL remains unchanged, wherein the input is an auxiliary variable, the variable dimension is 10, the output is F-CaO content, the dimension is 1, and the number of heads is 8.

[0032] Further, the TA-GCL uses an adjacency matrix to represent a topological graph A top , and the information of the field in the topological graph is aggregated through graph convolution;

[0033] The topological graph A top is used as the graph structure of graph convolution, and the A topThe middle element value is 0 and 1; when the elements of the adjacency matrix are binarized, the adjacency matrix is used to determine the connection between specific nodes in the graph, the nodes are input auxiliary variables, and one variable is one node; when the elements of the adjacency matrix are not binarized, the adjacency matrix reflects the relationship strength between nodes, and the nodes are all elements in the graph, including local spatiotemporal correlation centered on itself.

[0034] Further, the topological graph A top Further, a mask matrix M is included for determining edge strength, the mask matrix M is learnable, and the edge strength constitutes an adjacency matrix expressed as:

[0035] A M top +M

[0036] Further, the adjacency matrix and self-loop normalization are added to the adjacency matrix, the self-loop is in the form of a relationship matrix, and the adjacency matrix and self-loop normalization constitute a transformation matrix expressed as:

[0037]

[0038] Wherein, λ is a parameter for determining the strength of I, and I is a unit matrix.

[0039] Further, the expression of the graph convolution is:

[0040]

[0041] Wherein, the input data X G The size of the input data X PE is N*T, where N represents the number of nodes, T represents time, Q represents a learnable weight, l represents the number of layers of the model, and D is a degree matrix.

[0042] Further, the h-Head multi-head attention mechanism further includes using an embedding vector Emb to label the time information of the data, and the expression of the embedding vector is:

[0043]

[0044] X PE G +Emb

[0045] Wherein, pos=1,2,.T,i=1,2,.N,2i and 2i+1 represent even and odd respectively, and the PE is a position embedding mechanism.

[0046] Further, the h-Head multi-head attention mechanism is used to establish X PEThe long-term dependency in the time dimension also includes a residual connection, and the expression of the intermediate layer output data is:

[0047] X Att = h-Head(X PE , X PE , X PE ) + X PE

[0048] Further, the FC full connection layer maps the data to the target variable and outputs the data, and at the same time, the loss of the model is calculated by using a root mean square error, the loss is a regularization term, and the expression is:

[0049]

[0050] Wherein, T m is a further number of time, represents the real data, y i represents the output of the model, beta represents the regularization coefficient, H(.) represents the entropy, and is used to encourage structure discretization.

[0051] Further, the fourth step includes:

[0052] The cement operation process detection software fault diagnosis mode determines the running state of the production process according to the F-CaO content and the process data, and the running state includes stable, unstable and fault, wherein if it is determined to be unstable or even fault state, the cement operation process detection software fault positioning mode is entered to locate the fault variable, if it is a fault state, the running equipment also needs to be stopped, and the cement production process is fed back according to the determination result.

[0053] Another purpose of the present application is to provide a deep learning-based cement production tail gas monitoring system for implementing the deep learning-based cement production tail gas monitoring method, the deep learning-based cement production tail gas monitoring system comprising:

[0054] A data acquisition module is used for acquiring cement production process data;

[0055] A preprocessing module is used for preprocessing and regularizing the acquired data;

[0056] A TA-GCN module is used for screening data, training a prediction model, and realizing accurate prediction of the free calcium oxide content;

[0057] The cement operation process monitoring software includes a fault diagnosis mode and a fault positioning mode, and is used for judging whether the equipment fails and the production process is stable and efficient according to auxiliary variables and predicted variables, and positioning the fault.

[0058] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the application are analyzed from the following aspects:

[0059] The application provides a cement production process monitoring method based on a deep learning model, which realizes accurate prediction of a difficult-to-measure variable, i.e., free calcium oxide content, by using an easily-measured auxiliary variable.

[0060] According to the auxiliary variable and the predicted variable, it can be judged whether the equipment is malfunctioning and whether the production process is stable and efficient, so as to reduce the additional exhaust gas emission caused by incomplete combustion of raw materials.

[0061] The model disclosed by the application is not limited to cement exhaust emission, and can also be used for predicting other key variables of exhaust emission.

[0062] The cement production exhaust monitoring method based on deep learning disclosed by the application can assist workers in troubleshooting and reduce the time of incomplete combustion, so as to achieve the purpose of energy saving and emission reduction.

[0063] The expected income and commercial value of the technical solutions of the application after transformation are: taking the F-CaO concentration soft measurement model as an example, the accurate prediction of F-CaO concentration is realized in the cement production process, the emission of additional exhaust gas is effectively reduced, and the effect of energy saving and emission reduction is achieved. Once there is a sign of failure in the production process, the operator is reminded in time to reset the control parameters, so as to avoid the kiln from stopping due to failure or running for a long time, and reduce the loss in actual production.

[0064] The technical solutions of the application fill the technical gap in the industry at home and abroad: the traditional deep learning model is broken out of the dilemma of not being applied to industry due to the black box structure, and the accurate prediction of the difficult-to-measure variable, i.e., free calcium oxide content, is realized by using the easily-measured auxiliary variable, so as to determine whether it is in a non-stable or even failure state.

[0065] Whether the technical solutions of the application solve the technical problems that people have been eager to solve but have failed to achieve success: the carbon reduction space of the cement industry through the existing energy-saving technology to replace raw materials technology is very limited, and it is difficult to achieve the carbon reduction goal of the cement industry before the emergence of a revolutionary technology. The application aims to reduce or even eliminate unnecessary exhaust gas emission in the cement production process, so as to reduce the emission of carbon dioxide and other harmful exhaust gas in the cement industry, and achieve the goal of low-carbon and environmental protection industry.

[0066] Whether the technical solutions of the application overcome the technical bias: the traditional deep learning model is broken out of the dilemma of not being applied to industry due to the black box structure. BRIEF DESCRIPTION OF DRAWINGS

[0067] Figure 1is a cement production tail gas monitoring method based on deep learning provided by the embodiment of the present application;

[0068] Figure 2 is a TA-GCN model provided by the embodiment of the present application;

[0069] Figure 3 is a TA-GCL structure schematic diagram provided by the embodiment of the present application;

[0070] Figure 4 is a comparison diagram of F-CaO content prediction value and offline measured real value provided by the embodiment of the present application. DETAILED DESCRIPTION

[0071] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0072] In order to enable those skilled in the art to fully understand how the present application is specifically implemented, this part is an explanation and description of the embodiments of the technical scheme of the claims.

[0073] As shown in Figure 1 the cement production tail gas monitoring method based on deep learning provided by the embodiment of the present application comprises:

[0074] Step one, obtaining process detection data set of cement production;

[0075] Step two, preprocessing the data set;

[0076] Step three, outputting predicted F-Cao content data by using a time attention graph convolution network TA-GCN model;

[0077] Step four, judging equipment operating condition based on the F-Cao content data.

[0078] Further, the detection data set is collected by various sensors in the production equipment, and at the same time, auxiliary variables are used as inputs of the TA-GCN model, the auxiliary variables being feeding amount, kiln tail coal feeding amount, kiln head coal feeding amount, decomposing furnace outlet temperature, smoke chamber temperature, rotary kiln current, primary air fan frequency, carbon monoxide concentration, nitrogen oxide concentration, and decomposing furnace middle temperature.

[0079] The input variables selected for modeling by the TA-GCN model are respectively the feeding amount, the kiln tail coal feeding amount, the kiln head coal feeding amount, the decomposing furnace outlet temperature, the flue chamber temperature, the rotary kiln current, the primary air fan frequency, the carbon monoxide concentration, the nitrogen oxide concentration, and the decomposing furnace middle temperature, which are respectively denoted as V1, V2, V3, V4, V5, V6, V7, V8, V9, and V10. The upper half of the data in each month is used to train the model, and the lower half of the data in each month is used to test the effect of the software. Since the number of faults in actual cement production is insufficient to accurately train the software, simulated software is used to generate cement production fault data to train and test the software.

[0080] Further, the preprocessing is a regularization process, and the process is as follows:

[0081] The MinMaxScaler method is used to standardize the data, and the normalized data X scaled is:

[0082]

[0083] X scaled = X std *(max-min)+min

[0084] where X is the data to be normalized, X min is a row vector composed of the minimum values in each column of the data X, X max is a row vector composed of the maximum values in each column of the data X, max is the maximum value of the interval to be mapped, and min is the minimum value of the interval to be mapped. The interval is [0, 1].

[0085] As shown in Figure 2 , the time attention graph convolution network TA-GCN includes feature selection based on the maximum information coefficient MIC, a plurality of time attention graph convolution layers TA-GCL, and an FC full connection layer. The input data is first filtered, then the data is aggregated in a graph structure, and a plurality of TA-GCLs are stacked to establish a hierarchical model to capture complex spatial-temporal correlations, and finally the data is output through the FC full connection layer.

[0086] In particular, the TA-GCL aggregates data in the spatial dimension according to the topological graph and establishes long-term dependencies in the temporal dimension. Initially, the process data is summarized in a spatial dimension by GCL. In order to avoid the over-smoothing effect caused by too deep GCL, a residual connection is introduced. The input and output dimensions of each TA-GCL are uniformly set, and the number of heads of each TA-GCL remains unchanged to reduce hyperparameters.

[0087] Further, the maximum information coefficient MIC is a result of normalization of the maximum value of mutual information MI, which is obtained from the joint probability distribution p(J, U) of two consecutive random variables J and U, and represents how much the uncertainty of a random variable is reduced by introducing another random variable. The expression of the MI is:

[0088]

[0089] The expression of the MIC is:

[0090]

[0091] Wherein, B = T 0.55 , T represents the time step of data.

[0092] As Figure 3 shown, the TA-GCL includes a GCL and a multi-head attention mechanism MHA based on h-Head;

[0093] First, the data is aggregated in the spatial dimension according to the topological graph, and the long-term dependency relationship is established in the time dimension by using the multi-head attention mechanism MHA based on h-Head;

[0094] Secondly, the residual connection is introduced, the input and output dimensions of each TA-GCL are uniformly set, and the number of heads of each TA-GCL remains unchanged, wherein the input is an auxiliary variable, the variable dimension is 10, the output is the F-CaO content, the dimension is 1, and the number of heads is 8.

[0095] Further, the TA-GCL uses an adjacency matrix to represent the topological graph A top , and aggregates the information of the fields in the topological graph through graph convolution;

[0096] The topological graph A top is used as the graph structure of graph convolution, and the element value of the A top is 0 and 1; when the elements of the adjacency matrix are binarized, the adjacency matrix is used to determine the connection between specific nodes in the graph, and the node is an input auxiliary variable, one variable is one node; when the elements of the adjacency matrix are not binarized, the adjacency matrix reflects the relationship strength between nodes, and the node is all elements in the graph, including local spatiotemporal correlation centered on itself.

[0097] Further, the topological graph A top also includes a mask matrix M for determining edge strength, and the mask matrix M is learnable, and the expression of the adjacency matrix composed of the edge strength is:

[0098] A M = A top + M

[0099] Further, the adjacency matrix and the self-loop normalized are added to the adjacency matrix, the self-loop is in the form of a relationship matrix, and the expression of the transformation matrix composed of the adjacency matrix and the self-loop normalized is:

[0100]

[0101] Wherein, λ is a parameter for determining the intensity of I, I is a unit matrix.

[0102] Further, the expression of the graph convolution is:

[0103]

[0104] Wherein, the input data X G The size is N*T, wherein N represents the number of nodes, T represents time; Q represents a learnable weight, l represents the number of layers of the model, and D is a degree matrix.

[0105] Further, the h-Head multi-head attention mechanism further includes labeling the time information of the data by using the embedding vector Emb, and the expression of the embedding vector is:

[0106]

[0107] X PE =X G +Emb

[0108] Wherein, pos=1,2,.T,i=1,2,.N,2i and 2i+1 represent even and odd respectively, and the PE is a position embedding mechanism.

[0109] Further, the h-Head multi-head attention mechanism is used to establish X PE The long-term dependence in the time dimension also includes a residual connection, and the expression of the intermediate layer output data is:

[0110] X Att =h-Head(X PE ,X PE ,X PE )+X PE

[0111] Further, the FC full connection layer maps the data to the target variable, and outputs the data, and at the same time, the loss of the model is calculated by using the root mean square error, and the loss is a regularization term, and the expression is:

[0112]

[0113] Wherein, T m is a further number of time y represents real data, i y represents the output of the model, β represents the regularization coefficient, H(.) represents the entropy, and is used to encourage structure discretization.

[0114] Further, the illustrated step four includes:

[0115] The cement running process detection software fault diagnosis mode determines the running state of the production process according to the F-CaO content and process data, and the running state includes stable, unstable, and fault, wherein if it is determined as unstable or even fault state, the cement running process detection software fault positioning mode is entered to position the fault variable, if it is a fault state, the running equipment also needs to be stopped, and the cement production process is fed back and controlled according to the determination result.

[0116] In order to prove the creativity and technical value of the technical scheme of the application, this part is an application embodiment of the technical scheme of the claim on a specific product or related technology.

[0117] The cement production tail gas monitoring method based on deep learning provided by the application embodiment is applied to a computer device, the computer device includes a memory and a processor, the memory stores a computer program, and the computer program is executed by the processor to enable the processor to execute the steps of the cement production tail gas monitoring method based on deep learning.

[0118] The cement production tail gas monitoring method based on deep learning provided by the application embodiment is applied to an information data processing terminal, and the information data processing terminal is used to realize the cement production tail gas monitoring system based on deep learning.

[0119] It should be noted that the embodiments of the present application can be realized by hardware, software or a combination of software and hardware. The hardware part can be realized by special logic; the software part can be stored in the memory and executed by the appropriate instruction execution system, such as microprocessor or special designed hardware. Those skilled in the art can understand that the above-mentioned devices and methods can be realized by computer executable instructions and / or included in processor control code, such as carrier medium, such as magnetic disk, CD or DVD-ROM, programmable memory, such as read-only memory (firmware), or data carrier, such as optical or electronic signal carrier. The device of the present application and its modules can be realized by hardware circuit, such as ultra large scale integrated circuit or gate array, semiconductor, such as logic chip, transistor, or programmable hardware device, such as field programmable gate array, programmable logic device, etc. It can also be realized by software executed by various types of processors, or by a combination of the above-mentioned hardware circuit and software, such as firmware.

[0120] The embodiment of the present application has achieved some positive effects in research and development or use, and indeed has great advantages compared with the prior art, which is described below in combination with data, charts and the like of the experimental process.

[0121] Table 1 is the fault diagnosis accuracy rate, and Table 2 is the fault positioning accuracy rate, Figure 4 The soft measurement effect of the model. As can be seen from the chart, the present application accurately predicts the content of F-CaO, and exhibits superior performance in fault diagnosis and positioning.

[0122] Table 1 Fault diagnosis accuracy rate table

[0123]

[0124] Table 2 Fault positioning accuracy rate table

[0125]

[0126] The method of the present application uses a cement production tail gas monitoring method based on deep learning for the cement production process, and the software can determine whether it is in a non-stationary or even fault state, and give early warning and positioning for faults and fault variables, breaking through the dilemma that traditional deep learning models cannot be applied to industry due to black box structure. Taking the F-CaO concentration soft measurement model as an example, in the cement production process, accurate prediction of the F-CaO concentration is realized, effectively reducing the emission of additional tail gas, thereby achieving the effect of energy saving and emission reduction.

[0127] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any modification, equivalent replacement and improvement made by any person skilled in the art within the technical range disclosed by the present application should be covered within the protection scope of the present application.

Claims

1.A cement production tail gas monitoring method based on deep learning, characterized in that, The cement production tail gas monitoring method based on deep learning comprises: Step one, obtaining process detection data set of cement production; Step two, preprocessing the data set; Step three, outputting predicted F-Cao content data by using a time attention graph convolution network (TA-GCN) model; Step four, judging equipment operation condition based on the F-Cao content data; The TA-GCN model comprises feature selection based on maximum information coefficient (MIC), multiple time attention graph convolution layers (TA-GCLs), and a FC full connection layer; The process of the model is as follows: First, the input data is filtered, then the data is aggregated in a graph structure, multiple TA-GCLs are stacked to establish a hierarchical model to capture complex spatial-temporal correlation, and finally the data is output through the FC full connection layer; The MIC is a result of maximum information coefficient (MIC) normalization, the MI is obtained from the joint probability distribution p(J, U) of two continuous random variables J and U, and the expression of the MI is as follows: The expression of the MIC is as follows: where B = T 0.55 , T represents the time step of data; The TA-GCL comprises a graph convolution (GCL) and a multi-head attention mechanism (MHA) based on h-Head; First, the data is aggregated in the spatial dimension according to the topological graph, and the multi-head attention mechanism (MHA) based on h-Head is used to establish long-term dependence in the time dimension; Second, a residual connection is introduced, the input and output dimensions of each TA-GCL are uniformly set, and the number of heads of each TA-GCL remains unchanged, wherein the input variable dimension is 10, the output dimension is 1, and the number of heads is 8; The TA-GCL uses an adjacency matrix to represent a topology graph A top and aggregates information of fields in the topology graph by graph convolution GCL The topological graph A top is used as a graph structure for graph convolution, the A top The element values are 0 and 1; when the elements of the adjacency matrix are binarized, the adjacency matrix is used to determine the connection between specific nodes in the graph, the nodes are input auxiliary variables, and one variable is one node; when the elements of the adjacency matrix are not binarized, the adjacency matrix reflects the relationship strength between the nodes, and the nodes are all elements in the graph, including local spatiotemporal correlation centered on itself; The mask matrix M is learnable, and the expression of the adjacency matrix composed of edge strengths is as follows: A M = A top + M. 2.The cement production tail gas monitoring method based on deep learning according to claim 1, wherein, The adjacency matrix and the self-loop normalization are added to the adjacency matrix, the self-loop is in the form of a relation matrix, and the expression of the transformation matrix composed of the adjacency matrix and the self-loop normalization is as follows: Wherein, λ is a parameter for determining the strength of I, and I is an identity matrix. 3.The cement production tail gas monitoring method based on deep learning according to claim 1, wherein, The expression of the graph convolution is as follows: wherein the input data X G has a size of N x T, where N represents the number of nodes and T represents time; Q represents a learned weight, / represents the number of layers of the model, and D is a degree matrix. 4.The cement production tail gas monitoring method based on deep learning according to claim 1, wherein, The multi-head attention mechanism of h-Head further comprises using an embedding vector Emb to label the time information of the data, and the expression of the embedding vector is as follows: X PE = X G + Emb Wherein, pos = 1, 2,.T, i = 1, 2,.N, 2i and 2i+1 represent even and odd numbers respectively, and PE is a position embedding mechanism; The h-Head-based multi-head attention mechanism is used to establish X PE The long-term dependency in the time dimension also includes a residual connection, and the expression of the intermediate layer output data is: X Att = h - Head(X PE , X PE , X PE ) + X PE . 5.The cement production tail gas monitoring method based on deep learning according to claim 1, wherein, The FC full connection layer maps the data to a target variable and outputs the data, and at the same time, the loss of the model is calculated by using the root mean square error, and the loss is a regularization term, and the expression is as follows: where T m is the time further number, represents the real data, y i represents the output of the model, β represents the regularization coefficient, H(.) represents the entropy, used to encourage structure discretization; Step four comprises: The cement operation process detection software fault diagnosis mode determines the running state of the production process according to the F-CaO content and the process data, and the running state comprises stable, unstable, and fault, wherein if it is determined to be unstable or even a fault state, the cement operation process detection software fault positioning mode is entered to locate the fault variable, if it is a fault state, the running equipment needs to be stopped, and at the same time, the cement production process is fed back according to the determination result.

Images