JP2021143362A - Furnace condition learning method for blast furnace, furnace condition learning device, abnormality detection method, abnormality detection device and operation method - Google Patents

Abstract

To provide a furnace condition learning method and a furnace condition learning device for a blast furnace, even in the case where the causal relation between the furnace condition abnormality to be detected and measured data is uncertain or in the case where the time to a reception of influence by the furnace condition abnormality to be detected in the measured data is uncertain, capable of creating a neural network capable of detecting the furnace condition abnormality of the blast furnace at high precision.SOLUTION: A furnace condition learning method for a blast furnace comprises a first step where the image data of a raceway part of a blast furnace imaged in an imaging period including a period in which the furnace condition abnormality of the blast furnace is generated are learned by an unsupervised type neural network, a second step where, regarding each neuron composing the unsupervised type neural network after the learning, the correlation coefficient between the ignition value of the neuron and an index showing the furnace condition abnormality is calculated and a third step where, based on the correlation coefficient, the neuron used for detecting the furnace condition abnormality is extracted as a neuron for abnormality detection.SELECTED DRAWING: Figure 5

Description

The present invention relates to a furnace condition learning method, a furnace condition learning device, an abnormality detection method, an abnormality detection device, and an operation method of a blast furnace.

In recent years, a method of detecting a blast furnace condition abnormality by using a neural network learned by using data on a blast furnace condition known to be normal or abnormal has been proposed. Specifically, Patent Document 1 uses a neural network in which feature quantities extracted from a raceway image taken by an image pickup device whose focus has been adjusted in advance and sensory test results of an inspector are learned as teacher data. A method for determining the state of the blast furnace tuyere is described.

Japanese Patent No. 6350159

However, in the above method, when the causal relationship between the furnace condition abnormality to be detected and the measured data on the furnace condition of the blast furnace is uncertain, or when the measured data on the furnace condition of the blast furnace is affected by the furnace condition abnormality to be detected. If the time required to receive the furnace is unknown, it is not possible to create learning data for detecting the furnace condition abnormality, and it is not possible to accurately detect the furnace condition abnormality.

The present invention has been made in view of the above problems, and an object of the present invention is when the causal relationship between the furnace condition abnormality to be detected and the measured data is uncertain, or when the measured data is desired to detect the furnace condition abnormality. It is an object of the present invention to provide a blast furnace condition learning method and a furnace condition learning device capable of generating a neural network capable of accurately detecting a blast furnace condition abnormality even when the time until the influence of the above is unknown. Another object of the present invention is to provide a blast furnace abnormality detection method, an abnormality detection device, and an operation method capable of accurately detecting a blast furnace condition abnormality.

The method for learning the furnace condition of the blast furnace according to the present invention is the first step of learning the image data of the raceway portion of the blast furnace taken during the shooting period including the period when the furnace condition abnormality of the blast furnace occurs with an untrained neural network. Based on the second step of calculating the correlation coefficient between the firing value of the neuron and the exponent indicating the abnormal furnace condition for each neuron constituting the non-teachered neural network after learning, and the correlation coefficient. It is characterized by including a third step of extracting a neural network used for detecting an abnormality in the furnace condition as a neuron for detecting an abnormality.

In the above invention, the method for learning the furnace condition of the blast furnace according to the present invention is that in the first step, a set of a plurality of time-series image data of the raceway portion is learned as one voxel data by an untrained neural network. It is characterized by including steps to be performed.

In the method for learning the furnace condition of a blast furnace according to the present invention, in the above invention, in the first step, there is no set of a plurality of image data of the raceway portion taken by changing the focus of the imaging device as one voxel data. It is characterized by including a step of learning with a teacher-type neural network.

The furnace condition learning device for a blast furnace according to the present invention is a means for learning image data of a raceway portion of a blast furnace taken during a photographing period including a period in which an abnormality occurs in the furnace condition of the blast furnace by an untrained neural network. For each neuron constituting the untrained neural network after learning, a means for calculating the correlation coefficient between the firing value of the neuron and the index indicating the blast furnace condition abnormality, and the furnace condition abnormality based on the correlation coefficient. It is characterized by providing a means for extracting a neural network used for detecting an abnormality as an abnormality detection neural network.

The method for detecting an abnormality in a blast furnace according to the present invention includes a step of inputting image data of a raceway portion taken during operation of the blast furnace into an untrained neural network learned by the method for learning the furnace condition of the blast furnace according to the present invention. It is characterized by including a step of detecting a furnace condition abnormality of a blast furnace based on an ignition value of the abnormality detection neural network.

The blast furnace abnormality detection device according to the present invention is a means for inputting image data of a raceway portion taken during operation of the blast furnace into an untrained neural network learned by the blast furnace condition learning device according to the present invention. It is characterized by comprising means for detecting a furnace condition abnormality of a blast furnace based on the firing value of the abnormality detecting neuron.

The method for operating a blast furnace according to the present invention is characterized by including a step of operating the blast furnace while monitoring the abnormality of the blast furnace condition by using the method for detecting an abnormality of the blast furnace according to the present invention.

According to the blast furnace condition learning method and the blast furnace condition learning apparatus according to the present invention, it is possible to generate a neural network capable of accurately detecting the blast furnace condition abnormality. Further, according to the blast furnace abnormality detection method, the abnormality detection device, and the operation method according to the present invention, it is possible to accurately detect the blast furnace condition abnormality.

FIG. 1 is a schematic view showing a configuration example of a blast furnace to which an abnormality detection device for a blast furnace according to an embodiment of the present invention is applied. FIG. 2 is a schematic view showing a configuration example of a blast furnace to which the blast furnace abnormality detection device according to the embodiment of the present invention is applied. FIG. 3 is a schematic view showing an example of an image of a raceway taken by an imaging device. FIG. 4 is a block diagram showing a configuration of an abnormality detection device for a blast furnace according to an embodiment of the present invention. FIG. 5 is a flowchart showing the flow of the pre-learning process according to the embodiment of the present invention. FIG. 6 is a flowchart showing the flow of the abnormality detection process according to the embodiment of the present invention. FIG. 7 is a diagram showing an example of time-dependent changes in the aeration resistance index and the neuron firing value of the average of all tuyere.

Hereinafter, an abnormality detection device for a blast furnace, which is an embodiment of the present invention, will be described with reference to the drawings.

[Blast furnace configuration]
First, the configuration of the blast furnace to which the abnormality detection device for the blast furnace according to the embodiment of the present invention is applied will be described with reference to FIGS. 1 to 3.

FIG. 1 is a schematic view showing a configuration example of a blast furnace to which an abnormality detection device for a blast furnace according to an embodiment of the present invention is applied. As shown in FIG. 1, in order to blow hot air from a hot air furnace (not shown) into the blast furnace 1 inside the tuyere 2 of the blast furnace 1 to which the abnormality detection device for the blast furnace according to the embodiment of the present invention is applied. The blow pipe (blow pipe) 3 is connected, and the lance 4 is installed through the blow pipe 3. Fuels such as pulverized coal, oxygen, and city gas are blown into the blast furnace 1 from the lance 4. There is a combustion space called raceway 5 in the coke deposit layer in front of the hot air blowing direction of tuyere 2, and coke combustion and gasification (reduction of iron ore, that is, iron ore production) are mainly performed in this raceway 5. It is said. Further, as shown in FIG. 2, the blower pipe 3 is formed with a window 6 for monitoring the inside of the furnace for the operator to monitor the situation inside the blast furnace 1. Then, in the vicinity of the in-core monitoring window 6, an imaging device 7 for taking an image of the raceway 5 is installed through the in-core monitoring window 6.

FIG. 3 is a schematic view showing an example of an image of the raceway 5 taken by the image pickup apparatus 7. As shown in FIG. 3, in the image of the raceway 5 taken by the imaging device 7, the lance 4 and the raceway 5 are inside the circular image corresponding to the tip opening of the small tuyere 2a constituting the tuyere 2. The silhouette of is reflected. Approximately 40 tuyere 2s are installed in the circumferential direction of the blast furnace 1, and images of 40 raceways 5 in the circumferential direction of the blast furnace 1 are sequentially imaged by the imaging device 7. When the focus of the image pickup apparatus 7 is adjusted, the focus position changes inside the raceway 5. Therefore, if the focus is on the vicinity of the tuyere 2, it becomes easier to see the combustion state of the pulverized coal blown from the lance 4, and if the focus is shifted from the tuyere 2 to the inside of the raceway 5, the coke turning situation, etc. It becomes easier to see the internal state of the raceway 5. Further, by observing the time-series images of the raceway 5 taken by the image pickup apparatus 7, it is possible to confirm the change in the combustion state of the pulverized coal and the change in the turning state of the coke. By adjusting the focal position of the imaging device 7 in this way, the imaging from the tuyere 2 to the inside of the raceway 5 under a plurality of focal conditions and the time-series images thereof contain information related to the situation change in the blast furnace 1. include.

[Configuration of abnormality detection device]
Next, with reference to FIG. 4, the configuration of the abnormality detection device for the blast furnace according to the embodiment of the present invention will be described.

FIG. 4 is a block diagram showing a configuration of an abnormality detection device for a blast furnace according to an embodiment of the present invention. As shown in FIG. 4, the abnormality detection device for the blast furnace according to the embodiment of the present invention includes an information processing device 11, a display device 12, and an operating condition adjusting device 13.

The information processing device 11 is composed of an information processing device such as a computer, and functions as a pre-learning processing unit 11a and an abnormality detection unit 11b when an arithmetic processing unit such as a CPU inside the information processing device executes a computer program. The functions of the pre-learning processing unit 11a and the abnormality detection unit 11b will be described later.

The display device 12 is composed of a display device such as a liquid crystal display device, and displays various information according to a control signal from the information processing device 11.

The operating condition adjusting device 13 controls the operating conditions of the blast furnace 1 according to an operation input signal from the operator.

In the blast furnace abnormality detection device having such a configuration, the information processing device 11 executes the following pre-learning process and abnormality detection process to provide a neural network capable of accurately detecting the furnace condition abnormality of the blast furnace 1. At the same time as generating, the abnormal state of the blast furnace 1 is detected with high accuracy. Hereinafter, the operation of the information processing device 11 when executing the pre-learning process and the abnormality detection process will be described.

[Pre-learning process]
First, the operation of the information processing apparatus 11 when executing the pre-learning process will be described with reference to FIG.

An abnormality in the furnace condition of the blast furnace 1 represents, for example, a situation in which ventilation in the blast furnace 1 deteriorates, and is determined by a furnace condition index such as a ventilation resistance index. If any physical influence occurs in the vicinity of the tuyere 2 in response to such an abnormality in the furnace condition, the physical influence should be captured in the image of the raceway 5 taken by the image pickup apparatus 7. .. However, since the internal phenomenon of the blast furnace 1 has not been completely clarified, it is unclear whether the physical influence is observed in the vicinity of the tuyere 2. Also, even if it is known that there is a physical effect, it is not known how much the time when the physical effect is detected by the furnace condition index and the time when it is observed near the tuyere 2 are different. The furnace condition index cannot be associated with the image data of the raceway 5. This means that when machine learning is applied, it is not possible to create teacher data that determines whether the judgment of normal or abnormal is correct.

Therefore, the information processing device 11 executes a pre-learning process for adjusting the parameters of each neuron constituting the neural network in a situation where teacher data cannot be created. FIG. 5 is a flowchart showing the flow of the pre-learning process according to the embodiment of the present invention. As shown in FIG. 5, in the pre-learning process according to the embodiment of the present invention, first, the pre-learning process unit 11a has a tuyere 2 that is continuous in time including a time when the furnace condition is poor, which is photographed by the imaging device 7. Image data set of (step S1). Next, the pre-learning processing unit 11a unsupervisedly learns the neural network using the image data set of the tuyere 2 collected in the process of step S1 (step S2). Specifically, the pre-learning processing unit 11a takes the image data of the tuyere 2 as an input, and the output image data is reconstructed into the input image data set of the tuyere 2 so that the non-teacher type neural network is used. The parameters of each neural network that make up the network are optimized by iterative calculation. As a result, the firing value of each neuron is adjusted so that the value becomes larger than some image features of the image data set of tuyere 2 used for unsupervised learning, but what kind of image each neuron has. I don't know if it corresponds to the feature.

Therefore, the pre-learning processing unit 11a searches for a neuron whose firing value fluctuates in response to an abnormality in the furnace condition by calculating the correlation coefficient between the firing value of the neuron and the time-series data of the furnace condition index. Specifically, since the time difference (time delay) T between the time when the furnace condition abnormality is detected by the furnace condition index and the time observed near the tuyere 2 is unknown, first, the pre-learning processing unit 11a While waving the value of the time delay T in the firing value of the neuron, the time delay Tmax at which the absolute value of the correlation coefficient is maximized and the absolute value Rmax of the correlation coefficient at that time are obtained (step S3). Next, the pre-learning processing unit 11a compares the absolute value Rmax of the correlation coefficient obtained for each neuron among the neurons, and extracts the neuron having the maximum value Rmax2 of the absolute value Rmax of the correlation coefficient (at this time). The time delay is Tmax2) (step S4). Finally, the pre-learning processing unit 11a compares the furnace condition index with the firing value of the extracted neuron, and determines the threshold value S of the firing value of the neuron from which the period determined as the furnace condition abnormality can be extracted (step S5). ). As a result, a series of pre-learning processes is completed.

The maximum value Rmax2 of the correlation coefficient is a value calculated from the neurons having the highest correlation with the furnace condition index among the neurons unsupervised learned using the image data set of the tuyere 2. Therefore, if the maximum value Rmax2 of the correlation coefficient is sufficiently large, it can be determined that the influence of the furnace condition abnormality is reflected in the image data set of the tuyere 2. Further, if the time-delayed Tmax2 value is set to a condition in which the firing value of the neuron fluctuates earlier than the furnace condition index, in other words, the firing value of the neuron is set to the time Tmax2 with respect to the furnace condition index. If the time delay Tmax2 is greater than zero when the correlation coefficient is calculated with a shift, it is possible to predict the furnace condition abnormality in advance using the firing value of the neuron. As described above, in order to use the firing value of the neuron for the abnormality detection of the blast furnace 1, the firing value of the neuron needs to fluctuate before the furnace condition index. Therefore, when searching for the maximum value Rmax2 of the correlation coefficient, it is desirable to search under the condition that the time delay Tmax2 is larger than zero.

[Abnormality detection processing]
Next, with reference to FIG. 6, the flow of the abnormality detection process according to the embodiment of the present invention will be described.

FIG. 6 is a flowchart showing the flow of the abnormality detection process according to the embodiment of the present invention. As shown in FIG. 6, in the abnormality detection process according to the embodiment of the present invention, the image pickup apparatus 7 first captures the image data of the tuyere 2 (step S11). Next, the anomaly detection unit 11b inputs the image data of the tuyere 2 taken in the process of step S11 into the neural network trained by the pre-learning process, thereby determining the firing value of each neuron constituting the neural network. Calculate (step S12). Then, when the firing value of the neuron becomes equal to or higher than the predetermined threshold value S, the abnormality detection unit 11b determines that the furnace condition abnormality corresponding to the neuron will occur after the time Tmax, and notifies the operator to that effect. The warning information is output to the display device 12 (step S13). The operator adjusts the operating conditions of the blast furnace 1 by operating the operating condition adjusting device 13 in response to the warning information being displayed on the display device 12. According to such a process, the operating state of the blast furnace 1 can be quickly adjusted according to the warning information, so that the degree of abnormality in the furnace condition of the blast furnace 1 can be reduced.

〔Example〕
In this embodiment, an example in which the abnormality detection process is applied to the actual tuyere image data and the furnace condition abnormality will be described. Image data was collected by installing an imaging device in a blast furnace equipped with 40 tuyere in the circumferential direction. The image data set was obtained by extracting one sheet for each tuyere for one hour during the period of 11 days including the abnormal furnace condition. The image data was an RGB color image of 320 × 240 pixels. In addition, deterioration of ventilation was selected as the furnace condition abnormality, and the ventilation resistance index was used as the furnace condition index. In addition, as a neural network, an unsupervised 12-layer neural network described in the literature (Le et al., Building High-level Features Using Large Scale Unsupervised Learning, ICML2012) was adopted. Although this neural network is a locally connected neural network, unsupervised learning is possible using a fully connected neural network and image data such as CNN (Convolutional Neural Network) and GAN (Generative Adversarial Network). It may be a neural network.

Unsupervised learning of the neural network was performed by inputting the image data set according to the flowchart shown in FIG. Here, the calculation method of the correlation coefficient will be described. First, time-series images of each tuyere were input to the neural network after unsupervised learning, and the firing value of each neuron was calculated for each tuyere. Next, the firing value was normalized for each neuron. As the normalization method, min-max normalization using the maximum / minimum value and z-score normalization using the average value and the standard deviation are well known, but in this embodiment, the former is used. However, the latter may be used. The firing value data normalized for each neuron was averaged between tuyere to obtain the firing value of all tuyere averages. In this embodiment, the average firing value of all tuyere was obtained by averaging the output of firing value of neurons between tuyere, but the images of all 40 imaging devices were arranged side by side to form one image data. It may be used as the firing value of the neurons of the neural network learned by inputting the obtained data.

Time delay Tmax at which the absolute value of the correlation coefficient is maximized by shaking the time delay T with respect to the ignition value of all tuyere averages and the furnace condition index under the condition of 0 <time delay T <8 hours, and the phase relationship at that time. The absolute value Rmax of the number was calculated. Here, 8 hours is the average cycle time until the raw material put into the blast furnace becomes hot metal, and it is considered that there is no physical causal relationship for a longer time. Then, the absolute value Rmax of the correlation coefficient was compared between each neuron, and the neurons having the maximum value Rmax2 of the absolute value Rmax of the correlation coefficient were extracted. At this time, the maximum value Rmax2 of the absolute value Rmax of the correlation coefficient was 0.41, and the time delay Tmax2 at that time was 1 hour. In general, when the correlation coefficient is about 0.4, the correlation is not so large, but it is sufficient considering that not all the factors that caused poor ventilation are observed at the tuyere. It was decided to judge as.

FIG. 7 shows a plot showing the time variation of the aeration resistance index and the average neuron firing value of all tuyere. The neuron firing values are plotted one hour off in consideration of the time delay Tmax2. As shown in FIG. 7, when the ventilation resistance index is high, it is judged that the furnace condition is bad, and when the ventilation resistance index exceeds 1.1, the degree of the furnace condition abnormality is particularly high. In the latter half of the date 1/17, the ventilation resistance index rises sharply, and it can be seen that the ventilation is particularly poor. It has been shown that the neuron firing value rises one hour earlier than the aeration resistance index in exploration, but when the plots shown in FIG. 7 are qualitatively compared, it is shown that the neuron firing value rises to around 1.1 indicated by arrows R1 and R2. It can be seen that the neuron firing value is high several hours before that. Therefore, for example, if the threshold value S of the neuron firing value is set to 0.26, the blast furnace can be operated so as to suppress the furnace condition abnormality in advance, such as reducing the air flow amount before the furnace condition abnormality occurs. In this way, even if the causal relationship between the aeration of the detection target and the captured image is uncertain, the feature amount of the image is learned by the neurons by unsupervised learning, and the neuron that has a high correlation with the aeration is searched for in the next step. By doing so, an index that can be used for abnormality detection can be obtained. Further, by searching for the time delay T, it is possible to clarify the time until the captured image affects the ventilation.

Since one image is usually a cutout of the inside of the furnace at a certain moment, the information in the furnace obtained from the image is a change in spatial color that can be seen on the raceway. On the other hand, abnormal furnace conditions such as deterioration of ventilation in the furnace are information on changes over time. For this reason, it is possible to capture information on changes in color over time by integrating a plurality of images continuously captured in time series and treating them as one voxel data, so that the detection performance is improved. .. For example, if the image data in the above learning example is converted into voxel data by integrating 10 images (for 10 seconds) taken once per second, and the voxel data is collected every hour for 40 tuyere. good.

Further, if the focal point of the imaging device is fixed, only information on the depth of the raceway having a depth in the furnace direction can be obtained. For this reason, it is possible to capture information in the depth direction by integrating a plurality of images captured at a plurality of focal points and treating the image data as one voxel data, so that the detection performance is improved. For example, if the image data in the above learning example is made into voxel data in which the focus is adjusted in 10 steps and the 10 images captured by each are integrated, and the voxel data is collected every hour for 40 tuyere. good. Further, both a time-series image and an image with a changed focus may be adopted. In that case, for example, the image data in the above learning example is captured by 10 images (10 seconds) in which the focus is imaged once per second in each of 10 steps, and all the image data are integrated into voxel data. Forty voxels may be collected every hour.

Although the embodiment to which the invention made by the present inventors has been applied has been described above, the present invention is not limited by the description and the drawings which form a part of the disclosure of the present invention according to the present embodiment. That is, other embodiments, examples, operational techniques, and the like made by those skilled in the art based on the present embodiment are all included in the scope of the present invention.

1 Blast furnace 2 Tuft 3 Blower pipe (blow pipe)
4 Rance 5 Raceway 6 In-core monitoring window 7 Imaging device 11 Information processing device 11a Pre-learning processing unit 11b Abnormality detection unit 12 Display device 13 Operating condition adjustment device

Claims (7)

The first step of learning the image data of the raceway part of the blast furnace taken during the shooting period including the period when the blast furnace condition abnormality occurred with an untrained neural network,
For each neuron that constitutes an untrained neural network after learning, the second step of calculating the correlation coefficient between the firing value of the neuron and the exponent indicating the abnormal furnace condition, and
The third step of extracting the neuron used for detecting the furnace condition abnormality as the abnormality detection neuron based on the correlation coefficient, and
A method for learning the state of a blast furnace, which is characterized by including. The blast furnace according to claim 1, wherein the first step includes a step of learning a set of a plurality of time-series image data of the raceway portion as one voxel data by an untrained neural network. How to learn the furnace condition. The first step is characterized by including a step of learning a set of a plurality of image data of the raceway portion taken by changing the focus of the imaging device as one voxel data by an untrained neural network. The method for learning the furnace condition of a blast furnace according to claim 1 or 2. A means to learn the image data of the raceway part of the blast furnace taken during the shooting period including the period when the blast furnace condition abnormality occurred with an untrained neural network,
For each neuron that constitutes an untrained neural network after learning, a means for calculating the correlation coefficient between the firing value of the neuron and the index indicating the abnormal furnace condition, and
A means for extracting a neuron used for detecting the furnace condition abnormality based on the correlation coefficient as an abnormality detection neuron, and a means for extracting the abnormality.
A blast furnace condition learning device characterized by being equipped with. The step of inputting the image data of the raceway portion taken during the operation of the blast furnace into the non-teacher type neural network learned by the furnace condition learning method of the blast furnace according to any one of claims 1 to 3. ,
A step of detecting an abnormality in the blast furnace condition based on the firing value of the abnormality detection neuron, and
A method for detecting anomalies in a blast furnace, which comprises. A means for inputting image data of a raceway portion taken during operation of a blast furnace into an untrained neural network learned by the furnace condition learning device of the blast furnace according to claim 4.
A means for detecting an abnormality in the blast furnace condition based on the firing value of the abnormality detection neuron, and
Anomaly detection device for a blast furnace, characterized by being equipped with. A method for operating a blast furnace, which comprises a step of operating the blast furnace while monitoring the abnormality of the blast furnace condition using the method for detecting an abnormality in the blast furnace according to claim 5.

Images