Fault detection method for blast furnace steel production process based on MTF and Y-CNN

Abstract

The invention relates to a blast furnace steel production process fault detection technology, and aims to provide a blast furnace steel production process fault detection method based on MTF and Y-CNN. Comprising the following steps: an offline training stage: collecting sufficient related variable data, performing standardization processing, performing MTF conversion, and converting one-dimensional sequence data into two-dimensional image data; annotating the image and then inputting the image into a Y-CNN model for training; and in the online monitoring stage, variable data in the operation process of the blast furnace ironmaking system are collected in real time, sample data are processed, a graph is input into the trained Y-CNN model, and whether a fault exists in the blast furnace ironmaking production process or not and the fault type are determined according to an output diagnosis result. According to the invention, fault diagnosis is realized by using the Y-CNN model, so that the method has good robustness; faults with different feature sizes can be effectively detected, and the method can be well adapted to different working environments; the advantage that potential features can be learned by using multi-scale fusion and an attention mechanism is utilized, and the method can better adapt to a working environment.

Description

Technical Field

[0001] The present invention relates to the field of production process fault detection, and in particular to a blast furnace steel production process fault detection method based on MTF and Y-CNN. Background Art

[0002] The blast furnace production process refers to the process used to produce large-scale steel. The blast furnace is mainly composed of the furnace top, furnace body, tuyere area, hearth, and iron and slag outlets. Through continuous research and improvement in furnace design and operating parameter control, modern blast furnaces have not only improved production efficiency, but also reduced energy consumption and emissions, achieving a more environmentally friendly production method; automated control systems are widely used in contemporary blast furnace ironmaking processes, including real-time monitoring of furnace conditions, data analysis, and intelligent decision support systems to achieve optimal operating conditions. However, due to poor working conditions, various abnormal conditions such as difficult furnace conditions, collapsed materials, suspended materials, and furnace temperature heating / cooling often occur in the blast furnace ironmaking process. If they cannot be discovered and diagnosed in a timely and accurate manner, not only will the quality of molten iron deteriorate, but it may even threaten the safety of equipment and personnel. Therefore, fault diagnosis is crucial to ensure the safe operation of the blast furnace ironmaking process.

[0003] In recent years, with the increasing requirements for the quality of steel products, R&D personnel have proposed a variety of solutions for fault diagnosis in the blast furnace ironmaking process, trying to detect related abnormal conditions in a timely manner through fault monitoring and diagnosis, so as to make timely adjustments to ensure the quality of molten iron and equipment safety. For example, the industry commonly uses fault diagnosis methods such as support vector machine (SVM), random forest (RF), and deep convolutional neural network (DCNN), but they all have low accuracy due to their own defects.

[0004] Convolutional Neural Networks (CNN) is a deep feedforward neural network with local connections and weight sharing. It is one of the representative algorithms of deep learning and is good at processing images, especially image recognition and other related machine learning problems. For example, it has a significant improvement effect in various visual tasks such as image classification, target detection, and image segmentation. It is one of the most widely used models. Researchers have proposed a variety of research results for fault diagnosis based on convolutional neural networks. For example, the "structural damage identification method and device based on multi-channel MTF and CNN" proposed in patent document CN116522654A and the "gearbox online operation state abnormality identification method based on MTF-CNN technology" proposed in patent document CN118568540A combine Markov transfer fields and convolutional networks to perform multi-feature classification prediction. However, these technologies are prone to the defect of not being able to identify faults well due to the simple fault diagnosis model CNN. Summary of the invention

[0005] The technical problem to be solved by the present invention is to overcome the deficiencies in the prior art and provide a blast furnace steel production process fault detection method based on MTF and Y-CNN.

[0006] To solve the technical problem, the solution of the present invention is:

[0007] A method for fault detection in blast furnace steel production process based on MTF and Y-CNN is provided, which includes an offline training phase and an online monitoring phase; wherein:

[0008] The offline training phase includes:

[0009] (1.1) Collect sufficient data on the variables involved in the blast furnace ironmaking system under normal operation and fault conditions;

[0010] (1.2) After the data is standardized, MTF conversion is performed to convert the one-dimensional sequence data into two-dimensional image data;

[0011] (1.3) Annotate the obtained two-dimensional image and then input it into the Y-CNN model for training;

[0012] The online monitoring phase includes:

[0013] (2.1) real-time collection of variable data of the blast furnace ironmaking system during operation, and then processing the sample data with reference to the operation in step (1.2) in the offline training phase;

[0014] (2.2) Input the graph in step (2.1) into the trained Y-CNN model, and determine whether there is a fault in the blast furnace ironmaking production process and the type of fault based on the output diagnosis results.

[0015] As a preferred embodiment of the present invention, the variable data at least includes the following data during the operation of the blast furnace ironmaking system: 2 ,CO,CO 2 , oxygen-enriched flow rate, blast furnace top pressure, cold air pressure, total pressure difference, hot air pressure, resistance coefficient and actual coal injection amount.

[0016] As a preferred solution of the present invention, the operation of the MTF conversion is specifically as follows:

[0017] First, define the state space and discretize the values ​​in the one-dimensional time series by setting thresholds or dividing intervals;

[0018] Secondly, calculate the state transition probability and the transition probability between different states in the time series;

[0019] Finally, a Markov transition field is constructed and the state transition probability matrix P is converted into a Markov transition field.

[0020] As a preferred solution of the present invention, when converting the state transition probability matrix P into a Markov transition field, the following conversion formula is used:

[0021]

[0022] MTF[i][j]=P(s i ,s j )

[0023] Among them, s i is the state of the i-th position of the time series after normalization and discretization; x i is the i-th sequence point in time; max(X) refers to the maximum value of all data points in the time series; min(X) refers to the minimum value of all data points in the time series; Q is the number of divided state intervals; P[i][j] represents the number of states from state s i Transfer to state s j The probability of count(s i →s j ) indicates that from state s i The number of transitions to state sj; is the state i The sum of all the transition times of the outgoing state; MTF[i][j] is the element in MTF, representing the state s i To state s j The transfer intensity of P(s i ,s j ) indicates state s i To state s j The transition probability.

[0024] As a preferred embodiment of the present invention, the image annotation operation is specifically as follows: annotating the converted grayscale image according to the working conditions corresponding to when the variable data is generated; the annotation content includes whether it belongs to a normal working condition or a fault condition, and the corresponding fault type when the variable data is abnormal.

[0025] As a preferred solution of the present invention, the image annotation tool labelme is used to annotate the image.

[0026] As a preferred embodiment of the present invention, the Y-CNN model includes an input layer, a convolution layer, a pooling layer, a full connection and an output layer; wherein the four convolution layers are arranged in sequence, respectively used to realize 3*3 convolution, BN batch normalization, ReLU activation function and average pooling processing; after being processed by BN batch normalization and ReLU activation function, the results are first input into a hybrid attention module, respectively, for enhancing the attention to spatial attention and channel attention to strengthen difference discrimination; after upsampling and feature fusion processing of the outputs of the two hybrid attention modules, the data is sent to the output layer through the feature flattening module and two fully connected layers, and the results are finally output.

[0027] As a preferred solution of the present invention, the image size input to the Y-CNN model is 640*640, and the number of channels is 3; the feature sizes of the outputs of the four convolutional layers are 320*320, 160*160, 80*80 and 40*40, and the number of channels are 64, 128, 256 and 512, respectively; the feature size of the upsampling output is 80*80, and the number of channels is 1024; the feature size of the feature fusion module output is 80*80, and the number of channels is 1280.

[0028] As a preferred solution of the present invention, in the online monitoring stage, the variable data collected in real time during the operation of the blast furnace ironmaking system is converted into a grayscale image and input into a trained Y-CNN model. After calculation, the model outputs a result that directly shows whether there is a fault and the type of fault.

[0029] The present invention further provides a computer device, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor executes the aforementioned blast furnace steel production process fault detection method based on MTF and Y-CNN.

[0030] The present invention also provides a computer-readable storage medium, which stores computer instructions, and the computer instructions are used to enable the computer to execute the aforementioned blast furnace steel production process fault detection method based on MTF and Y-CNN.

[0031] Description of the invention principle:

[0032] MTF (Markov Transfer Field) is a model that describes the state transition in a Markov process. In multi-feature classification prediction, MTF can be used to model the transfer relationship between features, thereby capturing the correlation between features. By modeling this relationship, MTF can provide better feature representation and help improve classification performance. CNN (Convolutional Neural Network) is a deep learning model for image processing and pattern recognition. In multi-feature classification prediction, CNN can be used to extract features from input data. Through the combination of convolutional layers and pooling layers, CNN can automatically learn the spatial hierarchy of features and has the characteristics of translation invariance.

[0033] Considering the drawbacks of simply applying CNN in the prior art, the present invention proposes to combine MTF and Y-CNN technology and use it for fault diagnosis. Among them, Y-CNN is a high-precision CNN model with multi-scale fusion and attention mechanism. Compared with the traditional CNN model, Y-CNN has made improvements in the fusion of features at different scales, which has brought about an improvement in the overall performance of faults with features of different sizes.

[0034] The present invention uses various variables of the blast furnace steel production process as raw data, which are converted into the input of the Y-CNN model. The output of the Y-CNN model is the fault type related to the variable (for example, abnormal pressure leads to excessive pressure failure). The fault diagnosis method based on MTF and Y-CNN has good robustness, because the network of the model can extract potential features from a large amount of data, can process multiple variable data at the same time, and can adapt well in different working environments, achieving good fault diagnosis results. The working environment of the blast furnace steel production process is complex and changeable, and there are many process variables and quality variables, so the fault diagnosis method based on MTF and Y-CNN is an effective solution.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1. The present invention proposes a steel air compressor fault diagnosis algorithm based on MTF and Y-CNN. By converting one-dimensional time domain variable data into a two-dimensional image and using the Y-CNN model to realize fault diagnosis, the method has good robustness.

[0037] 2. Because the network of this model can use the attention mechanism senet to extract potential features from a large amount of data, it can process multiple variable data at the same time, and also integrate features of different sizes, so that faults of different feature sizes can be effectively detected. It can adapt well in different working environments and achieve good fault diagnosis effects.

[0038] 3. The working environment of the blast furnace steel production process is complex and changeable, with many process variables and quality variables. Although the traditional data-driven fault diagnosis method can diagnose faults to a certain extent, the results in judging the fault type are not ideal. The fault diagnosis method of the present invention takes advantage of the multi-scale fusion and attention mechanism to learn the potential features, can better adapt to the working environment, and is more flexible to use; and the model can directly output the type of fault, which is more intuitive than the traditional data-driven fault diagnosis method. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a flow chart of a specific implementation mode of the present invention.

[0040] Figure 2 It is a two-dimensional image schematic diagram of normal data in the blast furnace steel production process converted by MTF.

[0041] Figure 3 It is a two-dimensional image schematic diagram of the MTF conversion of hanging faults in the blast furnace steel production process.

[0042] Figure 4 It is a two-dimensional image schematic diagram of the hot blast furnace fault during the MTF conversion blast furnace steel production process.

[0043] Figure 5 This is the Y-CNN network structure diagram.

[0044] Figure 6 This is the training result diagram of the ordinary CNN model.

[0045] Figure 7 This is the training result diagram of the Y-CNN model. DETAILED DESCRIPTION

[0046] The specific implementation modes of the present invention are described in detail below with reference to the accompanying drawings.

[0047] Part I Implementation of the Invention

[0048] like Figure 1 As shown, the fault detection method for blast furnace steel production process based on MTF and Y-CNN in the present invention includes an offline training stage and an online monitoring stage; wherein,

[0049] The offline training phase includes:

[0050] (1.1) Collect sufficient data on the variables involved in the blast furnace ironmaking system under normal operation and fault conditions;

[0051] The variable data at least include the following data of the blast furnace ironmaking system during operation:2 ,CO,CO 2 , oxygen-enriched flow rate, blast furnace top pressure, cold air pressure, total pressure difference, hot air pressure, resistance coefficient and actual coal injection amount.

[0052] (1.2) After the data is standardized, MTF conversion is performed to convert the one-dimensional sequence data into two-dimensional image data;

[0053] The operation of the MTF conversion is as follows: first, define the state space, and discretize the values ​​in the one-dimensional time series by setting thresholds or dividing intervals; second, calculate the state transition probability, and calculate the transition probability between different states in the time series; finally, construct the Markov transition field, and use the following conversion formula to convert the state transition probability matrix P into the Markov transition field:

[0054]

[0055] MTF[i][j]=P(s i ,s j )

[0056] Among them, s i is the state of the i-th position of the time series after normalization and discretization; x i is the i-th sequence point in time; max(X) refers to the maximum value of all data points in the time series; min(X) refers to the minimum value of all data points in the time series; Q is the number of divided state intervals; P[i][j] represents the number of states from state s i Transfer to state s j The probability of count(s i →s j ) indicates that from state s i The number of transitions to state sj; is the state i The sum of all the transition times of the outgoing state; MTF[i][j] is the element in MTF, representing the state s i To state s j The transfer intensity of P(s i ,s j ) indicates state s i To state s j The transition probability.

[0057] (1.3) Annotate the obtained two-dimensional image and then input it into the Y-CNN model for training;

[0058] The operations of labeling the image using the image labelme are as follows: the converted grayscale image is labeled according to the working conditions corresponding to the generation of the variable data; the annotation content includes whether it belongs to the normal working condition or the fault condition, and the corresponding fault type when the variable data is abnormal.

[0059] The Y-CNN model includes input layer, convolution layer, pooling layer, full connection and output layer; among them, the four convolution layers are arranged in sequence, which are used to implement 3*3 convolution, BN batch normalization, ReLU activation function and average pooling respectively; after being processed by BN batch normalization and ReLU activation function, the results are first input into a hybrid attention module to enhance the attention to spatial attention and channel attention to strengthen difference discrimination; after upsampling and feature fusion processing of the outputs of the two hybrid attention modules, the data is sent to the output layer through the feature flattening module and two fully connected layers, and finally the results are output.

[0060] The online monitoring phase includes:

[0061] (2.1) real-time acquisition of variable data of the blast furnace ironmaking system during operation, and then processing the sample data with reference to the operation in step (1.2) in the offline training phase, and converting the real-time acquired variable data into a grayscale image;

[0062] (2.2) Input the graph in step (2.1) into the trained Y-CNN model. The model outputs the result after calculation, directly showing whether there is a fault and the type of fault.

[0063] Part II Contents of a Verification Experiment

[0064] The following simulation experiment is designed based on a real blast furnace ironmaking process (BFIP) data set to verify the effectiveness of the fault detection method for blast furnace steel production process based on MTF and Y-CNN. The blast furnace ironmaking process in this simulation experiment is divided into four main units: blast furnace body, blast furnace top, pulverized coal injection system and hot air system. The data set comes from an ironmaking enterprise in Liuzhou, Guangxi Zhuang Autonomous Region.

[0065] like Figure 1 The figure is a flow chart of the fault detection method of blast furnace steel production process based on MTF and Y-CNN of the present invention. The method comprises the following steps:

[0066] 1. Offline training phase

[0067] (1) Collect variable data and perform MTF conversion.

[0068] Collect variable data of blast furnace ironmaking process under normal conditions and fault conditions (H 2 ,CO,CO 2, oxygen-enriched flow rate, blast furnace top pressure, cold air pressure, total pressure difference, hot air pressure, resistance coefficient, actual coal injection amount and other variables). First, define the state space, and discretize the values ​​in the one-dimensional time series by setting thresholds or dividing intervals; secondly, calculate the state transition probability, and calculate the transition probability between different states in the time series; finally, construct the Markov transition field, and convert the state transition probability matrix P into MTF. The specific conversion formula is as described above.

[0069] (2) Image labeling. Label the converted grayscale image according to the corresponding working conditions, including normal working conditions and fault conditions. Use the labelme tool to label the data according to the corresponding fault when the variable data is abnormal.

[0070] (3) Train the Y-CNN model.

[0071] Build as Figure 3 The Y-CNN model shown in the figure shows the model training process in a specific example.

[0072] Input represents the input image, with a size of 640*640 and a channel number of 3. Conv1-conv4 all represent convolution operations, including 3*3 convolution, BN batch normalization, ReLU activation function and average pooling. The output feature sizes are 320*320, 160*160, 80*80 and 40*40, and the channel numbers are 64, 128, 256 and 512, respectively.

[0073] senet is a hybrid attention module that includes spatial attention and channel attention. Since the complexity of the MTF image varies due to data differences, there may be subtle changes, such as blurring of the edges of the MTF image. In order to improve the accuracy of fault diagnosis, the hybrid attention model is used to make the model pay more attention to key areas and identify subtle differences. Figure 2 The figure shows an example of MTF conversion diagram, which is an experimental result diagram on the blast furnace ironmaking process dataset.

[0074] concat is a feature fusion module. Given the different durations of faults, there are significant differences in the performance of MTF images, such as the shape of MTF images. The feature fusion module can fuse feature information at different levels to improve the accuracy of fault diagnosis. The feature map output by this module is 80*80 in size and has 1280 channels.

[0075] flat is the feature flattening module; fc1 and fc2 represent the first and second fully connected layers respectively; output represents the output, which is the specific fault category.

[0076] 2. Online monitoring stage

[0077] (1) Real-time collection of variable data of the blast furnace ironmaking system during operation.

[0078] (2) After a variable sample is converted into a grayscale image, it is input into the trained Y-CNN model. The output of the model will directly indicate whether a fault occurs and what type of fault it is.

[0079] 3. Comparative experimental data and result analysis

[0080] from Figure 6 It can be seen that the training results obtained by combining MTF with ordinary CNN have an accuracy rate of less than 50%, while Figure 7 The training results obtained by combining MTF and Y-CNN can be obtained in Figure 1, and its accuracy is close to 60%. It can be seen that the combination of MTF and Y-CNN proposed in the present invention is very effective.

Claims

1. A fault detection method for blast furnace steel production process based on MTF and Y-CNN, characterized in that: The method includes an offline training phase and an online monitoring phase; wherein, The offline training phase includes: (1.1) Collect sufficient data on the variables involved in the blast furnace ironmaking system under normal operation and fault conditions; (1.2) After the data is standardized, MTF conversion is performed to convert the one-dimensional sequence data into two-dimensional image data; (1.3) Annotate the obtained two-dimensional image and then input it into the Y-CNN model for training; The online monitoring phase includes: (2.1) real-time collection of variable data of the blast furnace ironmaking system during operation, and then processing the sample data with reference to the operation in step (1.2) in the offline training phase; (2.2) Input the graph in step (2.1) into the trained Y-CNN model, and determine whether there is a fault in the blast furnace ironmaking production process and the type of fault based on the output diagnosis results.

2. The method according to claim 1, characterized in that: The variable data include at least the following data of the blast furnace ironmaking system during operation: H2, CO, CO2, oxygen-enriched flow, blast furnace top pressure, cold air pressure, total pressure difference, hot air pressure, resistance coefficient and actual coal injection amount.

3. The method according to claim 1, characterized in that The operation of the MTF conversion is as follows: First, define the state space and discretize the values ​​in the one-dimensional time series by setting thresholds or dividing intervals; Secondly, calculate the state transition probability and the transition probability between different states in the time series; Finally, a Markov transition field is constructed and the state transition probability matrix P is converted into a Markov transition field.

4. The method according to claim 3, characterized in that When converting the state transition probability matrix P to the Markov transition field, the following conversion formula is used: MTF[i][j]=P(s i ,s j ) Among them, s i is the state of the i-th position of the time series after normalization and discretization; x i is the i-th sequence point in time; max(X) refers to the maximum value of all data points in the time series; min(X) refers to the minimum value of all data points in the time series; Q is the number of divided state intervals; P[i][j] represents the number of states from state s i Transfer to state s j The probability of count(s i →s j ) indicates that from state s i The number of transitions to state sj; is the state i The sum of all the transition times of the outgoing state; MTF[i][j] is the element in MTF, representing the state s i To state s j The transfer intensity; P(s i ,s j ) indicates state s i To state s j The transition probability.

5. The method according to claim 1, characterized in that The image annotation operation is specifically as follows: annotate the converted grayscale image according to the working condition corresponding to when the variable data is generated; the annotation content includes whether it belongs to a normal working condition or a fault condition, and the corresponding fault type when the variable data is abnormal.

6. The method according to claim 1, characterized in that Use the image annotation tool labelme to annotate the image.

7. The method according to claim 1, characterized in that The Y-CNN model includes an input layer, a convolution layer, a pooling layer, a full connection and an output layer; wherein the four convolution layers are arranged in sequence, respectively used to implement 3*3 convolution, BN batch normalization, ReLU activation function and average pooling processing; after being processed by BN batch normalization and ReLU activation function, the results are first input into a hybrid attention module, respectively, to enhance the attention to spatial attention and channel attention to strengthen difference discrimination; after upsampling and feature fusion processing of the outputs of the two hybrid attention modules, the data are sent to the output layer through the feature flattening module and two fully connected layers, and the results are finally output.

8. The method according to claim 7, characterized in that The image size input to the Y-CNN model is 640*640, and the number of channels is 3; the feature sizes of the outputs of the four convolutional layers are 320*320, 160*160, 80*80 and 40*40, and the number of channels are 64, 128, 256, and 512 respectively; the feature size of the upsampling output is 80*80, and the number of channels is 1024; the feature size of the feature fusion module output is 80*80, and the number of channels is 1280.

9. The method according to claim 1, characterized in that: During the online monitoring stage, the variable data collected in real time during the operation of the blast furnace ironmaking system is converted into grayscale images and input into the trained Y-CNN model. After calculation, the output result of the model directly shows whether there is a fault and the type of fault.

10. A computer device, characterized in that: include: At least one processor, and a memory in communication with the at least one processor, wherein the memory stores instructions executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor executes the blast furnace steel production process fault detection method based on MTF and Y-CNN as described in any one of claims 1 to 9.

11. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable the computer to execute the blast furnace steel production process fault detection method based on MTF and Y-CNN as described in any one of claims 1 to 9.

Images