Early fault diagnosis method and device for circulating fluidized bed based on multi-view attention

Abstract

The present invention discloses a circulating fluidized bed early fault diagnosis method and device based on multi-view attention, the method steps include: step S01. constructing a CFB unit digital twin system, and mapping the CFB unit digital twin system with the real-time operation status data of the CFB unit; step S02. collecting multi-source operation data of the CFB digital twin system during operation and performing data fusion to obtain CFB fusion twin data; step S03. inputting the CFB fusion twin data into the multi-view attention network module to extract multi-scale fault features based on the attention mechanism; step S04. reconstructing samples according to the extracted multi-scale fault features and using the reconstructed samples to calculate the SPE index, and judging the fault state of the current CFB unit system according to the SPE index. The present invention has the advantages of simple implementation method, low cost, high diagnostic efficiency and accuracy, strong real-time performance and flexibility.

Description

Technical Field

[0001] The present invention relates to the technical field of coal power equipment fault diagnosis, and in particular to a circulating fluidized bed (CFB) early fault diagnosis method and device based on multi-view attention. Background Art

[0002] Coal gasification technology is the process of converting coal into clean fuel gas. In fact, through high-temperature gasification reaction, the organic matter in coal is converted into synthesis gas, which is a mixture of carbon monoxide and hydrogen. It can effectively reduce the emission of pollutants generated by coal combustion and improve energy efficiency. The core unit of coal gasification technology is the fluidized bed, among which the circulating fluidized bed (CFB) has the advantages of strong heat and mass transfer, uniform temperature distribution, large equipment production capacity, uniform mixing of solid particles, wide range of adjustment of the residence time of solid particles in the bed, and easy large-scale, and is currently widely used.

[0003] The CFB unit has high requirements for operational safety. Any minor fault may cause unplanned shutdown of the unit, and in serious cases may cause safety accidents. Therefore, it is necessary to monitor the faults of the CFB unit in real time. However, due to the diverse equipment structures of CFB units, large-scale, complex and difficult-to-control combustion processes, and many influencing factors and complex operating environments, more and more fault data need to be monitored. A large amount of fault data increases the difficulty of fault diagnosis, making it difficult to judge faults in a timely and accurate manner. Therefore, it is very important to conduct digital monitoring and intelligent fault prediction of CFB units.

[0004] However, in the prior art, the digital monitoring and intelligent fault prediction methods of equipment usually extract certain fault features of the equipment and use traditional fault diagnosis and identification methods (such as machine learning models) to determine the fault status of the equipment based on the fault features. This type of method has the following problems:

[0005] 1) Fault warning methods and models are usually relatively independent of physical entities and do not form a close connection, which leads to the inability to effectively combine mechanism prior knowledge and the insufficient generalization performance of fault diagnosis models, making it difficult to directly apply them to guide actual operation and maintenance activities of CFB units;

[0006] 2) Traditional fault diagnosis and identification methods can usually only realize fault diagnosis after the fault occurs, and it is difficult to realize the diagnosis of early faults of CFB units. In addition, the structure of CFB units is complex, and a single fault feature cannot accurately characterize the fault state of CFB units. A large number of fault features will increase the complexity of diagnosis and reduce diagnostic efficiency. When applied to complex systems, it is actually difficult to achieve fast and precise fault diagnosis, and it is not suitable for intelligent operation and maintenance of CFB units. Summary of the invention

[0007] The technical problem to be solved by the present invention is: in response to the technical problems existing in the prior art, the present invention provides a circulating fluidized bed early fault diagnosis method and device based on multi-perspective attention, which has a simple implementation method, low cost, high diagnostic efficiency and accuracy, and strong real-time and flexibility.

[0008] In order to solve the above technical problems, the technical solution proposed by the present invention is:

[0009] A circulating fluidized bed early fault diagnosis method based on multi-view attention, comprising the following steps:

[0010] Step S01. Build a digital twin system of a CFB unit, and map the digital twin system of the CFB unit with the real-time operating status data of the CFB unit, wherein the operating status data includes status information of the equipment in a normal state, status information of the equipment in an early fault state, and status information of the equipment in a fault state;

[0011] Step S02. Collect multi-source operation data of the CFB digital twin system during operation and perform data fusion to obtain CFB fused twin data, where the multi-source operation data includes real-time operation data, equipment performance parameters, and operation scheduling configuration parameters;

[0012] Step S03. Input the CFB fusion twin data into the multi-view attention network module to extract multi-scale fault features based on the attention mechanism, and the multi-scale fault features include macroscale, mesoscale, and microscale; in the multi-view attention network module, by inputting global particle concentration information, macroscale fault features, mesoscale fault features, and microscale fault features are extracted through the macroscale channel, mesoscale channel, and microscale channel using average pooling and maximum pooling, respectively, and each channel uses a different channel attention weight, wherein the channel attention weight of the macroscale channel is less than the channel attention weight of the mesoscale channel, and the channel attention weight of the mesoscale channel is less than the attention weight of the microscale channel;

[0013] Step S04. Reconstruct samples according to the extracted multi-scale fault features and use the reconstructed samples to calculate the SPE (Squared Prediction Error) index, and judge the fault state of the current CFB unit system according to the extracted multi-scale fault features and the SPE index, wherein it is judged whether the current CFB unit has an early fault according to the micro-scale fault features, the meso-scale fault features and the SPE index, and it is judged whether the current CFB unit has a mid-term fault according to the meso-scale fault features, the macro-scale fault features and the SPE index.

[0014] Furthermore, the process of building a digital twin system for the CFB unit also includes: configuring multiple fault types, and adjusting the ratio of coal powder to gasification agent to simulate the formation process of faults including coking inside the CFB and blockage of the return pipe, setting high-pressure inlet gas to increase the collision between particles to simulate the early fault formation process, and obtaining fault information features at various scales as a fault data set. The fault types include any of the following: blockage of the uniform air plate holes, the installation position of the uniform air plate exceeding a preset threshold, the CFB inlet angle being less than a preset threshold, and the CFB lining falling off.

[0015] Furthermore, in the multi-view attention network module, the calculation expression of the macro-scale fault feature extracted by the macro-scale channel based on the attention mechanism is:

[0016]

[0017] in, represents the macro-scale fault characteristics, represents the macro-scale attention weight, represents the mesoscale attention weight, represents the micro-scale attention weight, represents the attention function, denote the average channel and maximum pooling of macro-scale channels, respectively. represents the global particle concentration information, represents the block size used for scaling, Represents a convolutional layer used to reduce the channel size.

[0018] Furthermore, in the multi-view attention network module, the calculation expression for extracting the mesoscale fault features based on the attention mechanism of the mesoscale channel is:

[0019]

[0020] in, represents the mesoscale fault characteristics, represents the attention function, represents the mesoscale attention weight, represents the macro-scale attention weight, denote the average channel and maximum pooling of the mesoscale channel, respectively. represents the global particle concentration information, represents the convolutional layer used to reduce the channel size, Indicates the block size used for scaling.

[0021] Furthermore, in the multi-view attention network module, the calculation expression of the microscale fault feature extracted by the microscale channel based on the attention mechanism is:

[0022]

[0023] in, represents the microscale fault characteristics, represents the attention function, represents the micro-scale attention weight, denote the average channel and maximum pooling of micro-scale channels, respectively. represents the global particle concentration information, represents the convolutional layer used to reduce the channel size, Indicates the block size used for scaling.

[0024] Furthermore, in step S04, fault reconstruction is performed and the SPE index is calculated according to the following expression:

[0025]

[0026]

[0027]

[0028]

[0029]

[0030] in, represents the extracted multi-scale fault features, for The reconstructed sample vector of The fault amplitude portion that represents the output value deviating from the normal characteristics, The fault frequency portion that represents the output value deviating from the normal characteristics, is the superimposed L2 norm of fault amplitude and frequency, They are macro-scale fault characteristics, meso-scale fault characteristics and micro-scale fault characteristics. Macro-scale fault characteristics The macro-scale spatial feature vector obtained after feature vectorization is Mesoscale fault characteristics The mesoscale space feature vector obtained after feature vectorization is Microscale fault characteristics The micro-scale spatial feature vector obtained after feature vectorization is for No. i diagonal elements, is the gain matrix, is the identity matrix, yes No. i diagonal elements.

[0031] Further, in step S04, judging the fault state of the current CFB unit system according to the extracted multi-scale fault features and SPE index includes:

[0032] Define the objective function for:

[0033]

[0034] in, Indicates fault characteristics Or reconstruct the sample vector , for The control limit of is the Hotelling statistic, for The control limit of is a symmetric and positive definite matrix;

[0035] Using micro-scale spatial eigenvectors , mesoscale space feature vector Calculate the corresponding control limits , according to the control limit Calculate the corresponding , , and determine whether If yes, it is determined that the current CFB unit has an early fault, otherwise it is determined that there is no early fault, where the control limit Calculated according to the following expression:

[0036]

[0037]

[0038] Among them, P belongs to , represents the standard normal distribution, It represents the confidence level;

[0039] Using mesoscale spatial eigenvectors , macro-scale spatial feature vector Calculate the corresponding control limits , according to the control limit Calculate the corresponding , , and determine whether If yes, it is determined that the current CFB unit has a mid-term fault, otherwise it is determined that there is no mid-term fault, where the control limit Calculated according to the following expression:

[0040] ;

[0041] .

[0042] A circulating fluidized bed early fault diagnosis device based on multi-view attention, comprising:

[0043] A twin system construction module is used to construct a digital twin system of a CFB unit and map the digital twin system of the CFB unit with the real-time operation status data of the CFB unit, wherein the operation status data includes the status information of the equipment in a normal state, the status information of the equipment in an early fault state, and the status information of the equipment in a fault state;

[0044] The data fusion module is used to collect multi-source operation data of the CFB digital twin system during operation and perform data fusion to obtain CFB fused twin data. The multi-source operation data includes real-time operation data, equipment performance parameters, and operation scheduling configuration parameters.

[0045] A multi-scale fault feature extraction module is used to input the CFB fusion twin data into a multi-view attention network module to extract multi-scale fault features based on an attention mechanism, wherein the multi-scale fault features include macroscale, mesoscale, and microscale; in the multi-view attention network module, by inputting global particle concentration information, macroscale fault features, mesoscale fault features, and microscale fault features are extracted through macroscale channels, mesoscale channels, and microscale channels using average pooling and maximum pooling, respectively, and each channel uses a different channel attention weight, wherein the channel attention weight of the macroscale channel is less than the channel attention weight of the mesoscale channel, and the channel attention weight of the mesoscale channel is less than the attention weight of the microscale channel;

[0046] The fault status diagnosis module is used to reconstruct samples according to the extracted multi-scale fault features and calculate the SPE index using the reconstructed samples, and judge the fault status of the current CFB unit system according to the extracted multi-scale fault features and the SPE index, wherein the current CFB unit is judged whether there is an early fault according to the micro-scale fault features, the meso-scale fault features and the SPE index, and the current CFB unit is judged whether there is a mid-term fault according to the meso-scale fault features, the macro-scale fault features and the SPE index.

[0047] A computer device comprises a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to execute the computer program to perform the above method.

[0048] A computer-readable storage medium storing a computer program, wherein the computer program implements the above method when executed by a processor.

[0049] Compared with the prior art, the advantages of the present invention are: the present invention establishes a digital twin model of the fluidized bed based on the digital twin technology to simulate the operating status and performance parameters of the equipment, collects, analyzes and models real-time monitoring data of the fluidized bed equipment and system based on the digital twin technology, and combines the multi-perspective attention mechanism to effectively extract multi-scale fault characteristics of macroscale, mesoscale and microscale, and fully explore potential fault characteristics from different scales. Finally, the fault feature method is reconstructed by combining the SPE indicator to determine whether there is an early fault. It can realize comprehensive perception of intelligent fault diagnosis and prediction, establish a close connection between physical entities, effectively combine the mechanism prior knowledge, and improve the accuracy of early fault prediction. It can realize real-time, dynamic and accurate intelligent CFB unit early fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a schematic diagram of the implementation flow of the circulating fluidized bed early fault diagnosis method based on multi-view attention in this embodiment.

[0051] Figure 2 It is a schematic diagram of the detailed process principle of realizing early fault diagnosis of a circulating fluidized bed in this embodiment.

[0052] Figure 3 It is a schematic diagram of the architecture principle of the data fusion system constructed in this embodiment.

[0053] Figure 4 It is a schematic diagram of the principle flow of the multi-view attention module in this embodiment to realize the extraction of fault features of each scale based on attention blocks of different scales.

[0054] Figure 5 It is a schematic diagram of the principle of implementing fault diagnosis in this embodiment. DETAILED DESCRIPTION

[0055] The present invention is further described below in conjunction with the accompanying drawings and specific preferred embodiments, but the protection scope of the present invention is not limited thereby.

[0056] like Figure 1 , 2 As shown, the steps of the circulating fluidized bed early fault diagnosis method based on multi-view attention in this embodiment include:

[0057] Step S01. Build a digital twin system of the CFB unit, and map the digital twin system of the CFB unit with the real-time operating status data of the CFB unit. The operating status data includes status information of the equipment in normal state, status information of the equipment in early fault state, and status information of the equipment in fault state.

[0058] In this embodiment, a CFB unit system twin is constructed based on digital twin technology, and the CFB unit mechanism fault model is integrated into the CFB unit virtual entity, such as Figure 2 As shown in the figure, the architecture of the entire CFB unit system is designed according to the structural composition of the physical entity of the CFB unit system, and the virtual fault entity of the CFB unit system is constructed based on the Unity simulation platform, including a three-dimensional model, a physical model, a rule model, and a particle circulation model, etc. The rule model is a model that represents the flow rules of particles in the CFB when the particle operation simulation setting is performed on the CFB, and the particle circulation model is a model that represents the suspension and dense phase flow of particles when the particle operation simulation setting is performed on the CFB; the scene construction process includes building a workshop model, configuring the unit layout, model materials, etc., and scene rendering, etc. During the operation of the twin system, the real-time data of the three-dimensional model is mapped to the twin in real time. The real-time data includes the status information of the equipment in normal state, the status information of the equipment in early fault state, and the status information of the equipment in fault state. Then, the above information can be used to construct a fault feature set, and the key data of the CFB unit can also be visualized.

[0059] This embodiment builds a virtual twin model of the CFB unit system, which can break through the limitations of physical conditions and carry out a series of activities such as simulation, prediction, monitoring, optimization, control, and verification to achieve continuous and immediate response and upgrade optimization. It can facilitate the acquisition of multidimensional data sets under complex working conditions, realize data sharing throughout the entire life cycle, and solve the complexity and uncontrollability of the complex CFB unit system due to its internal combustion process. At the same time, it can closely link fault warnings and models with physical entities, and can effectively combine mechanism prior knowledge to improve the generalization performance of fault diagnosis models.

[0060] In this embodiment, the process of building a digital twin system of the CFB unit also includes: configuring multiple fault types, and adjusting the ratio of coal powder to gasification agent to simulate the formation process of faults including coking inside the CFB and blockage of the return pipe, setting high-pressure inlet gas to increase the collision between particles to simulate the early fault formation process, and obtaining fault information characteristics at various scales. The fault types include blockage of the uniform air plate hole, the installation position of the uniform air plate exceeding the preset threshold (that is, the installation position of the uniform air plate is too high), the CFB inlet angle is less than the preset threshold (that is, the CFB inlet angle is too small), and the CFB liner falls off.

[0061] In a specific application embodiment, the detailed steps of building a CFB unit digital twin system are as follows:

[0062] Step S101, based on the 3DMAX three-dimensional model, the three-dimensional preliminary modeling of the CFB unit equipment scene is completed by combining polygonal modeling, two-dimensional to three-dimensional and patch modeling to obtain a virtual power plant model.

[0063] Step S102, import the virtual power plant model obtained in step S101 into Unity software to perform mapping operations on some of the required models. After the mapping is completed, write a control program to control the movement of the model to realize all the movement processes by performing basic translation, rotation and scaling operations on the three-dimensional model.

[0064] Step S103, integrating the CFB unit mechanism fault model into the CFB unit virtual entity, and developing faults such as blockage of the air leveling plate hole, too high installation position of the air leveling plate, too small CFB inlet angle, CFB liner falling off, internal coking, etc.

[0065] Specifically, faulty parts are designed in advance, such as blockage of the air uniform plate holes, too high installation position of the air uniform plate, too small CFB inlet angle, detachment of the CFB liner, etc. The ratio of coal powder to gasification agent is adjusted, such as continuously increasing the proportion of coal powder to simulate the formation process of a series of faults such as coking inside the CFB and blockage of the return pipe. High-pressure inlet gas is set to increase the collision between particles and break them up to simulate the early fault formation process, and the fault information characteristics of each scale are obtained to form a fault data set.

[0066] In the digital twin virtual space, the activities of the entire life particle cycle are simulated, monitored, optimized and verified to achieve data sharing throughout the entire life cycle. During the operation of the CFB digital twin system, real-time data such as normal equipment status information, early equipment failure status information, and equipment failure status information are output as data sets.

[0067] Step S02: Collect multi-source data of the CFB unit system during production and operation and perform data fusion to obtain CFB fusion twin data.

[0068] This embodiment combines the operation data of the coal gasification CFB unit with the multi-source heterogeneous data of the digital twin CFB unit, and performs consistency verification on the CFB unit data based on the real CFB unit operation data to ensure the validity of the data and reduce the amount of data storage.

[0069] In a specific application embodiment, the detailed steps of implementing CFB unit data consistency check and data fusion are as follows:

[0070] Step S201: construct a data fusion system, such as Figure 3As shown in the figure, it is divided into four layers, namely data application layer, data processing layer, data storage layer and data perception layer. The data perception layer collects data from conveyor belts, collectors, PLC (programmable controller), RTU (remote control terminal) and CFB and other devices, and stores them in a real-time database. The real-time database can support continuous query of real-time data and real-time index positioning. The data processing layer can perform data fusion, elimination of redundant data, data preprocessing, true value processing and denoising on the data in the real-time database. The data application layer uses the data processed by the data processing layer to realize real-time monitoring, real-time data query and display, and data decision-making.

[0071] Step S202: pre-process the data for multi-source processing.

[0072] The production and operation data of CFB units include multiple sources and storage locations, including real-time data, reports, equipment performance parameters, operation scheduling configuration parameters, and staff work records. The real-time data set of CFB units has the following characteristics: (1) There are a large number of sensor measurement points in the CFB unit system, and the amount of data is huge; (2) The thermal system data has complex correlations and usually presents a nonlinear relationship, resulting in data complexity; (3) Due to the differences in the operating environment of CFB units, various noises and abnormal points are prone to occur, making it difficult to ensure data quality and accuracy. In this embodiment, in view of the above data characteristic (1), consistency processing and abnormal data elimination are used to perform consistency processing on multi-source data; in view of data characteristics (2) and (3), multi-source data fusion and normalized storage are adopted, and data fusion and normalized storage are performed on multi-source redundantly stored data to integrate the data into a unified data center for subsequent analysis and application.

[0073] Step S203, CFB unit operation data consistency detection: Before building the digital twin data center, first verify the consistency of the data to eliminate abnormal data caused by the harsh environment in which the CFB unit operates.

[0074] Specifically, the distribution graph method can be used to perform consistency checks and abnormal data according to the actual situation of the CFB unit and the requirements of high reliability and ease of implementation of the scenario.

[0075] Step S204, CFB unit operation data level fusion based on adaptive weighting: a multi-source data fusion method of a neural network with adaptive weighting is used to achieve the fusion of multi-source data to obtain CFB fusion twin data.

[0076] Specifically, weights can also be introduced w To mark the measurement accuracy, specific data processing can be achieved according to different accuracies. The size of the weight represents the importance of the corresponding measurement data.

[0077] Step S03. Input the CFB fusion twin data obtained in step S02 into the multi-view attention network module to extract multi-scale fault features based on the attention mechanism. The multi-scale fault features include macroscale, mesoscale, and microscale. In the multi-view attention network module, by inputting the global particle concentration information, the macroscale fault features, mesoscale fault features, and microscale fault features are extracted respectively through the macroscale channel, mesoscale channel, and microscale channel using average pooling and maximum pooling. Each channel uses a different channel attention weight, wherein the channel attention weight of the macroscale channel is smaller than the channel attention weight of the mesoscale channel, and the channel attention weight of the mesoscale channel is smaller than the attention weight of the microscale channel.

[0078] In the early fault state, the fault characteristics are not obvious and the diagnosis is difficult. This embodiment integrates the digital twin technology with the attention mechanism model, uses multi-scale channels to extract fault characteristics of different scales, and configures different channel attention weights for each channel based on the channel attention mechanism. The configuration makes the channel attention weights of the microscale channel, mesoscale channel, and macroscale channel decrease in sequence, which can make the focus on microscale features in the feature extraction process, followed by mesoscale features. It can not only effectively realize the extraction of features of different scales, but also improve the extraction accuracy of micro-scale features. Then, the extracted multi-scale fault features can be used to diagnose the early and mid-term states of the fault, so as to combine digital twins with deep learning methods to realize timely and precise CFB fault diagnosis and prediction.

[0079] In this embodiment, a set of weight distributions can be first learned autonomously based on the network. The weight distribution represents the degree. The ones with large weights are important feature information that the network is easy to pay attention to, and the ones with small weights are some hidden feature information. The network can be used to pay attention to the multi-scale fault information in the circulating fluidized bed through the multi-view attention mechanism. The multi-scale information of the circulating fluidized bed can be divided into three scales, macroscale, mesoscale and microscale. Among them, the macroscale fault features in the bed can be easily noticed at first glance, which can affect the global pressure trend in the entire bed, while the mesoscale fault features and microscale features are weakened in turn. Specifically, a set of weight information can be obtained through network autonomous learning, and it can be fused with the original CFB fault features. The fused feature information can reflect the feature details that the network needs to focus on.

[0080] In this embodiment, the solid phase diagram of different positions of the flow field of the lifting section in the three-dimensional gas-solid fluidized bed within a period of time is first obtained, and the calculation expression is as follows:

[0081] (1)

[0082] The input contains T The internal position in time is arrive Solid phase flow field diagram.

[0083] The particle concentration information of the lifting section sequence is defined as the ratio of the particle number concentration at the measurement position to the average particle number concentration in the bed under the reference condition. The particle concentration X formula is as follows:

[0084] (2)

[0085] in, It is the particle size concentration information at each location in the CFB bed, which is used to measure the spatial distribution of particles. is the inlet particle velocity, is the exit particle velocity.

[0086] This embodiment designs a multi-view attention mechanism module composed of macro-scale, meso-scale, micro-scale features and attention channels for fault features at each scale, where macro-scale, meso-scale and micro-scale represent fault signal information. The attention mechanism can automatically adjust the weight according to the importance of the input features, thereby focusing on key information. The construction of the multi-view attention mechanism module can be expressed by the following mathematical model:

[0087] (3)

[0088] in, Represents particle characteristics at a macroscale and can influence global pressure trends throughout the bed; Represents the particle characteristics at the mesoscale, which is located between macroscale particles and microscale particles and is the particle agglomeration of the local flow of CFB; Represents the particle characteristics at the microscale, which are the characteristics of tiny particles broken in the collision among high-speed flowing particles in CFB; Attention mechanism, which can automatically adjust the weights according to the importance of input features to focus on key information; Represents a nonlinear function or model that combines the attention mechanism and features at different scales to produce the final output. Function Specifically, it is a complex deep learning model that captures key information through the attention mechanism. Represented by the function The resulting mapping output.

[0089] In a specific application embodiment, Figure 4 As shown in the figure, the detailed execution steps of the multi-view attention module to extract fault features at each scale based on attention blocks at different scales are as follows:

[0090] (1) Macro-scale fault feature extraction.

[0091] Macro-scale features are the manifestation of information on a global or larger scale. Therefore, in the attention mechanism, macro-scale feature attention can be achieved by considering the global particle concentration information. x Specifically, we use global average pooling (GAP) and global max pooling (GMP) to extract global information from the input feature map. Assuming that the input feature map is the particle concentration map at each position of the lift tube, its size is C×H×W, where C represents the number of channels, H and W represent the height and width respectively, and the global output context macro-scale information is represented by , whose size is C × H × W .

[0092] Represent macro-scale features and define attention functions function that accepts a macroscale feature 、 Mesoscale Attention Weights and represents the micro-scale attention weights As input, it outputs an attention weight vector with the same dimension as the feature ， The size is C×1×1, that is, the macro-scale attention weight vector is obtained . The attention weight vector 、 Mesoscale Attention Weights and represents the micro-scale attention weights After multiplication, they form the channel attention weight of the macro-scale channel, and then are combined with the macro-scale feature Perform element-wise multiplication to obtain the weighted macro-scale features , whose size is C×H×W, and the weighted operation formula can be expressed as:

[0093] (4)

[0094] Based on the above weighted macro-scale features The macro-scale features can be obtained. The calculation expression of the macro-scale channel in the multi-view attention network module to extract the macro-scale features based on the attention mechanism can be expressed as:

[0095] (5)

[0096] in, represents the macro-scale fault characteristics, represents the macro-scale attention weight, , represents the mesoscale attention weight, represents the micro-scale attention weight, Represents the attention function, the global attention of the self-attention mechanism outputs macro-scale granular fault characteristics , represents the block size used for scaling, denote the average channel and maximum pooling of macro-scale channels, respectively. represents the global particle concentration information, represents a convolutional layer used to reduce the channel size, e.g. is a convolutional layer that reduces the channel size from 2 to 1. Attention function For the self-attention mechanism, the output passes through the activation function sigmoid to ensure that the element values ​​of the weight vector are between 0 and 1. , , They are values ​​between 0 and 1.

[0097] (2) Mesoscale fault feature extraction.

[0098] Specifically, the pooling layer method can be used to extract the input features M f Extract local information. Assume that the obtained local context feature map is Flocal, with a size of Cc×Hh×Ww, where Cc represents the number of channels of the local feature, and Hh and Ww represent the height and width of the local feature, respectively. Fuse the local context feature map with the original input feature map, such as element-by-element multiplication. Assume that the fused feature map is Ffused, with a size of Cc×Hh×Ww (the dimension can be adjusted through a convolutional layer or a fully connected layer). Input the fused feature map Ffused into the fully connected layer, and apply the activation function to calculate the attention weight. Assume that the obtained attention weight is , whose size is Cc×Hh×Ww, and the mesoscale attention weight , micro-scale attention weight After multiplication, it is used as the channel attention weight of the mesoscale channel, and then multiplied element by element with the fused feature map Ffused to obtain the weighted feature map, which can be expressed as:

[0099] (6)

[0100] The feature map represents the local sequence features in each block. Using the convolutional layer to design local attention can reduce the complexity of the model. By using a smaller block size and a larger kernel size, the convolutional layer can fully represent the intermediate sequence features. A depth convolution layer with a kernel size of Cc / 2 - 1 is used, and after the depth convolution layer, the mesoscale attention weights are estimated by concatenating channel average and max pooling. 。

[0101] In a specific application embodiment, in the multi-view attention network module, the computational expression for the mesoscale channel to extract mesoscale fault features based on the attention mechanism can be expressed as:

[0102] (7)

[0103] Where, represents the mesoscale fault feature, represents the attention function, represents the mesoscale attention weight, represents the macroscale attention weight, respectively represent the average channel and max pooling of the mesoscale channel, represents the convolution layer for reducing the channel size, represents the block size for scaling.

[0104] According to the above mechanism, the first block is input , then, the mesoscale attention weight and the macroscale attention weight are determined, the channel attention weight of the mesoscale channel is obtained, and the mesoscale particle fault feature is output based on the self-attention mechanism global attention therein 。

[0105] (3)Microscale fault feature extraction

[0106] In this embodiment, for the microscale feature attention mechanism, it focuses more on capturing details and local information. By using the microscale attention weight matrix as the channel attention weight of the microscale channel and multiplying it with each channel of the feature map to achieve channel microscale attention weighting. Since the channel attention weight of the microscale channel is greater than that of the mesoscale and macroscale channels, more attention can be concentrated on the extraction of microscale fault features, enabling effective diagnosis of early faults.

[0107] In a specific application embodiment, in the multi-view attention network module, the computational expression for the microscale channel to extract microscale fault features based on the attention mechanism can be expressed as:

[0108] (8)

[0109] (9)

[0110] Where, represents the micro-scale attention weight, denote the average channel and maximum pooling of micro-scale channels, respectively. represents the convolutional layer used to reduce the channel size, represents the block size used for scaling, Represents microscale features.

[0111] The multi-view attention network module of this embodiment uses channel attention to characterize the features of each scale. Each pool output passes through a shared linear layer. The core of the channel attention mechanism is to assign a weight value to each channel to highlight the channels that contribute more to the task and suppress irrelevant channels.

[0112] Step S04: reconstruct samples according to the extracted multi-scale fault features and calculate the SPE index, and judge the fault status of the current CFB unit system according to the SPE index.

[0113] like Figure 5 As shown in the figure, during the failure process of the CFB unit, it first passes through the early fault point and then gradually evolves into the fault point. The early fault threshold is smaller than the threshold of the fault point. When fault maintenance is performed in time in the early fault stage, the occurrence of faults can be effectively avoided.

[0114] The SPE indicator can detect industrial processes with multiple variables, showing the deviation of the measured value at a certain moment relative to the steady-state operation, and reflecting the information hidden in the original data. When the CFB unit state is abnormal, the SPE value will increase significantly.

[0115] Step S401: reconstruct samples of the extracted multi-scale fault features and calculate the SPE index.

[0116] The core task of sample reconstruction is to apply corrections to the erroneous data and Estimate the normal value without the influence of fault , thereby eliminating the influence of fault amplitude and spectrum and improving the accuracy of estimation. Since early faults do not have obvious fault signs, fault characteristics are often weak and difficult to identify and discover. The same mode often has the characteristics of overlapping fault characteristics, that is, there will be additive faults in the measurement, so the data space is decomposed into a macro-scale data space and meso- and micro-scale data spaces , , the following fault model can be used to reconstruct the extracted multi-scale fault features:

[0117] (10)

[0118] (11)

[0119] (12)

[0120] (13)

[0121] in, represents the extracted multi-scale fault features, for The reconstructed sample vector of The fault amplitude portion that represents the output value deviating from the normal characteristics, The fault frequency portion that represents the output value deviating from the normal characteristics, is the superposition L2 norm of fault amplitude and frequency, They are macro-scale fault characteristics, meso-scale fault characteristics and micro-scale fault characteristics. Macro-scale fault characteristics The macro-scale spatial feature vector obtained after feature vectorization is Mesoscale fault characteristics The mesoscale space feature vector obtained after feature vectorization is Microscale fault characteristics The micro-scale spatial feature vector obtained after feature vectorization is for No. i diagonal elements, is the gain matrix, is the identity matrix, yes No. i diagonal elements.

[0122] Based on the above fault model, the SPE function can be reconstructed and the SPE index can be calculated according to the following formula:

[0123] (14)

[0124] in, yes No. i diagonal elements.

[0125] Step S402. Define the objective function for fault identification and analysis based on SPE indicators for:

[0126] (15)

[0127] (16)

[0128] in, express Or reconstruct the sample vector , for The control limit of is the Hotelling statistic, for The control limit, is a symmetric and positive definite matrix.

[0129] Step S403. Using micro-scale spatial feature vector , mesoscale space feature vector Calculate the corresponding control limits , according to the control limit Calculate the corresponding , , and determine whether If yes, it is determined that the current CFB unit has an early fault, otherwise it is determined that there is no early fault.

[0130] Specifically, the control limit It can be calculated according to the following expression:

[0131] (17)

[0132] (18)

[0133] Among them, P belongs to , represents a standard normal distribution (mean 0, variance 1), It represents the confidence level, that is, the upper 0.05 quantile of the standard normal distribution and the critical value at the 95% confidence level.

[0134] Step S404. Use the mesoscale space feature vector , macro-scale spatial feature vector Calculate the corresponding control limits , according to the control limit Calculate the corresponding , , and determine whether If so, it is determined that the current CFB unit has a mid-term fault, otherwise it is determined that there is no mid-term fault.

[0135] Specifically, the control limit It can be calculated according to the following expression:

[0136] (19)

[0137] (20)

[0138] The complexity of multidimensional data under complex working conditions makes the corresponding relationship between faults and symptoms complex, and the faults show diversity and transmission, which affects the decision of faults. How to process complex multidimensional data is the key to ensuring data validity and fault prediction accuracy. The present invention establishes a digital twin model of a fluidized bed based on digital twin technology to simulate the operating status and performance parameters of the equipment. Based on digital twin technology, the real-time monitoring data of the fluidized bed equipment and system are collected, analyzed and modeled. At the same time, the multi-view attention mechanism is combined to effectively extract the multi-scale fault characteristics of macroscale, mesoscale and microscale, and the potential fault characteristics are fully excavated from different scales. Finally, the fault feature method is combined with the SPE index to reconstruct the fault feature method to determine whether there is an early fault, which can realize comprehensive perception of intelligent fault diagnosis and prediction, establish a close connection between physical entities, effectively combine the mechanism prior knowledge, and improve the accuracy of early fault prediction. In addition, combined with data twin technology, the fault prediction results can also be traceable and reasonable, and the real-time, dynamic, accurate and scientific intelligent fault diagnosis can be realized, and the prediction accuracy of CFB unit system faults can be improved to ensure the stable operation of the unit.

[0139] This embodiment further provides a circulating fluidized bed early fault diagnosis device based on multi-view attention, comprising:

[0140] A twin system construction module is used to construct a digital twin system of a CFB unit and map the digital twin system of the CFB unit with the real-time operation status data of the CFB unit, wherein the operation status data includes the status information of the equipment in a normal state, the status information of the equipment in an early fault state, and the status information of the equipment in a fault state;

[0141] The data fusion module is used to collect multi-source operation data of the CFB digital twin system during operation and perform data fusion to obtain CFB fused twin data. The multi-source operation data includes real-time operation data, equipment performance parameters, and operation scheduling configuration parameters.

[0142] A multi-scale fault feature extraction module is used to input the CFB fusion twin data into a multi-view attention network module to extract multi-scale fault features based on an attention mechanism, wherein the multi-scale fault features include macroscale, mesoscale, and microscale; in the multi-view attention network module, by inputting global particle concentration information, macroscale fault features, mesoscale fault features, and microscale fault features are extracted through macroscale channels, mesoscale channels, and microscale channels using average pooling and maximum pooling, respectively, and each channel uses a different channel attention weight, wherein the channel attention weight of the macroscale channel is less than the channel attention weight of the mesoscale channel, and the channel attention weight of the mesoscale channel is less than the attention weight of the microscale channel;

[0143] The fault status diagnosis module is used to reconstruct samples according to the extracted multi-scale fault features and calculate the SPE index using the reconstructed samples, and judge the fault status of the current CFB unit system according to the extracted multi-scale fault features and the SPE index, wherein the current CFB unit is judged whether there is an early fault according to the micro-scale fault features, the meso-scale fault features and the SPE index, and the current CFB unit is judged whether there is a mid-term fault according to the meso-scale fault features, the macro-scale fault features and the SPE index.

[0144] The circulating fluidized bed early fault diagnosis device based on multi-view attention in this embodiment corresponds one to one with the above-mentioned circulating fluidized bed early fault diagnosis method based on multi-view attention, and will not be described one by one here.

[0145] The present invention can be applied to provide key protection for various large-scale equipment in production, improve equipment stability and production efficiency, reduce the waste of various raw coals, reduce the loss of equipment consumables and reduce harmful gas emissions, and can also be used to achieve the optimal design of CFB unit structure and predict and deduce faults.

[0146] This embodiment further provides a computer device, including a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to execute the computer program to perform the above method.

[0147] It is understandable that the above method of this embodiment can be executed by a single device, such as a computer or server, etc., and can also be applied to a distributed scenario and completed by multiple devices in cooperation with each other. In the case of a distributed scenario, one of the multiple devices can only execute one or more steps in the above method of this embodiment, and multiple devices interact to complete the above method. The processor can be implemented in the form of a general-purpose CPU, a microprocessor, an application-specific integrated circuit, or one or more integrated circuits, etc., for executing related programs to implement the above method of this embodiment. The memory can be implemented in the form of a read-only memory ROM, a random access memory RAM, a static storage device, and a dynamic storage device. The memory can store an operating system and other applications. When the above method of this embodiment is implemented by software or firmware, the relevant program code is stored in the memory and called and executed by the processor.

[0148] This embodiment further provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the above method is implemented.

[0149] Those skilled in the art should understand that the above-mentioned embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application can take the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the process Figure 1 A process or multiple processes and / or boxes Figure 1 These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer-readable memory produce a product including an instruction device, which implements the functions specified in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide for implementing the process in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0150] The above is only a preferred embodiment of the present invention, and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment, it is not intended to limit the present invention. Therefore, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the scope of protection of the technical solution of the present invention.

Claims

1. A method for early fault diagnosis of a circulating fluidized bed based on multi-view attention, characterized in that the steps include: Step S01. Build a digital twin system of a CFB unit, and map the digital twin system of the CFB unit with the real-time operating status data of the CFB unit, wherein the operating status data includes status information of the equipment in a normal state, status information of the equipment in an early fault state, and status information of the equipment in a fault state; Step S02. Collect multi-source operation data of the CFB digital twin system during operation and perform data fusion to obtain CFB fused twin data, where the multi-source operation data includes real-time operation data, equipment performance parameters, and operation scheduling configuration parameters; Step S03. Input the CFB fusion twin data into the multi-view attention network module to extract multi-scale fault features based on the attention mechanism, wherein the multi-scale fault features include macro-scale, meso-scale, and micro-scale; In the multi-view attention network module, by inputting global particle concentration information, macroscale fault features, mesoscale fault features and microscale fault features are correspondingly extracted through macroscale channels, mesoscale channels and microscale channels using average pooling and maximum pooling, and each channel uses a different channel attention weight, wherein the channel attention weight of the macroscale channel is less than the channel attention weight of the mesoscale channel, and the channel attention weight of the mesoscale channel is less than the attention weight of the microscale channel; Step S04. Reconstruct samples according to the extracted multi-scale fault features and use the reconstructed samples to calculate the SPE index, and judge the fault state of the current CFB unit system according to the extracted multi-scale fault features and the SPE index, wherein it is judged whether the current CFB unit has an early fault according to the micro-scale fault features, the meso-scale fault features and the SPE index, and it is judged whether the current CFB unit has a mid-term fault according to the meso-scale fault features, the macro-scale fault features and the SPE index.

2. The method for early fault diagnosis of a circulating fluidized bed based on multi-view attention according to claim 1 is characterized in that: The process of building a digital twin system for the CFB unit also includes: configuring multiple fault types, and adjusting the ratio of coal powder to gasification agent to simulate the formation process of faults including coking inside the CFB and blockage of the return pipe, setting high-pressure inlet gas to increase the collision between particles to simulate the early fault formation process, and obtaining fault information features at various scales as a fault data set. The fault types include any of the following: blockage of the uniform air plate hole, the installation position of the uniform air plate exceeding the preset threshold, the CFB inlet angle being less than the preset threshold, and the CFB lining falling off.

3. The method for early fault diagnosis of a circulating fluidized bed based on multi-view attention according to claim 1 is characterized in that: In the multi-view attention network module, the calculation expression of the macro-scale fault feature extracted by the macro-scale channel based on the attention mechanism is: in, represents the macro-scale fault characteristics, represents the macro-scale attention weight, represents the mesoscale attention weight, represents the micro-scale attention weight, represents the attention function, denote the average channel and maximum pooling of macro-scale channels, respectively. represents the global particle concentration information, represents the block size used for scaling, Represents a convolutional layer used to reduce the channel size.

4. The method for early fault diagnosis of a circulating fluidized bed based on multi-view attention according to claim 1 is characterized in that: In the multi-view attention network module, the calculation expression of the mesoscale fault feature extracted by the mesoscale channel based on the attention mechanism is: in, represents the mesoscale fault characteristics, represents the attention function, represents the mesoscale attention weight, represents the macro-scale attention weight, denote the average channel and maximum pooling of the mesoscale channel, respectively. represents the global particle concentration information, represents the convolutional layer used to reduce the channel size, Indicates the block size used for scaling.

5. The method for early fault diagnosis of a circulating fluidized bed based on multi-view attention according to claim 1 is characterized in that: In the multi-view attention network module, the calculation expression of the micro-scale fault features extracted by the micro-scale channel based on the attention mechanism is: in, represents the microscale fault characteristics, represents the attention function, represents the micro-scale attention weight, denote the average channel and maximum pooling of micro-scale channels, respectively. represents the global particle concentration information, represents the convolutional layer used to reduce the channel size, Indicates the block size used for scaling.

6. The method for early fault diagnosis of a circulating fluidized bed based on multi-view attention according to any one of claims 1 to 5, characterized in that: In step S04, fault reconstruction is performed and the SPE index is calculated according to the following expression: in, represents the extracted multi-scale fault features, for The reconstructed sample vector of The fault amplitude portion that represents the output value deviating from the normal characteristics, The fault frequency portion that represents the output value deviating from the normal characteristics, is the superposition L2 norm of fault amplitude and frequency, They are macro-scale fault characteristics, meso-scale fault characteristics and micro-scale fault characteristics. Macro-scale fault characteristics The macro-scale spatial feature vector obtained after feature vectorization is Mesoscale fault characteristics The mesoscale space feature vector obtained after feature vectorization is Microscale fault characteristics The micro-scale spatial feature vector obtained after feature vectorization is for No. i diagonal elements, is the gain matrix, is the identity matrix, yes No. i diagonal elements.

7. The method for early fault diagnosis of a circulating fluidized bed based on multi-view attention according to claim 6 is characterized in that: In step S04, judging the fault status of the current CFB unit system according to the extracted multi-scale fault features and SPE index includes: Define the objective function for: in, Indicates fault characteristics Or reconstruct the sample vector , for The control limit of is the Hotelling statistic, for The control limit of is a symmetric and positive definite matrix; Using micro-scale spatial eigenvectors , mesoscale space feature vector Calculate the corresponding control limits , according to the control limit Calculate the corresponding , , and determine whether If yes, it is determined that the current CFB unit has an early fault, otherwise it is determined that there is no early fault, where the control limit Calculated according to the following expression: Among them, P belongs to , represents the standard normal distribution, It represents the confidence level; Using mesoscale spatial eigenvectors , macro-scale spatial feature vector Calculate the corresponding control limits , according to the control limit Calculate the corresponding , , and determine whether If yes, it is determined that the current CFB unit has a mid-term fault, otherwise it is determined that there is no mid-term fault, where the control limit Calculated according to the following expression: ； 。 8. A circulating fluidized bed early fault diagnosis device based on multi-view attention, characterized in that: include: A twin system construction module is used to construct a digital twin system of a CFB unit and map the digital twin system of the CFB unit with the real-time operation status data of the CFB unit, wherein the operation status data includes the status information of the equipment in a normal state, the status information of the equipment in an early fault state, and the status information of the equipment in a fault state; The data fusion module is used to collect multi-source operation data of the CFB digital twin system during operation and perform data fusion to obtain CFB fused twin data. The multi-source operation data includes real-time operation data, equipment performance parameters, and operation scheduling configuration parameters. A multi-scale fault feature extraction module, used to input the CFB fusion twin data into a multi-view attention network module to extract multi-scale fault features based on an attention mechanism, wherein the multi-scale fault features include macro-scale, meso-scale, and micro-scale; In the multi-view attention network module, by inputting global particle concentration information, macroscale fault features, mesoscale fault features and microscale fault features are correspondingly extracted through macroscale channels, mesoscale channels and microscale channels using average pooling and maximum pooling, and each channel uses a different channel attention weight, wherein the channel attention weight of the macroscale channel is less than the channel attention weight of the mesoscale channel, and the channel attention weight of the mesoscale channel is less than the attention weight of the microscale channel; The fault status diagnosis module is used to reconstruct samples according to the extracted multi-scale fault features and calculate the SPE index using the reconstructed samples, and judge the fault status of the current CFB unit system according to the extracted multi-scale fault features and the SPE index, wherein the current CFB unit is judged whether there is an early fault according to the micro-scale fault features, the meso-scale fault features and the SPE index, and the current CFB unit is judged whether there is a mid-term fault according to the meso-scale fault features, the macro-scale fault features and the SPE index.

9. A computer device comprising a processor and a memory, wherein the memory is used to store a computer program, wherein: The processor is configured to execute the computer program to perform the method according to 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 a processor, the method according to any one of claims 1 to 7 is implemented.

Images