TW202144588A - Blast-furnace furnace condition learning method, furnace condition learning device, abnormality detection method, abnormality detection device, and operation method - Google Patents

Abstract

This blast-furnace furnace condition learning method comprises: a first step in which image data of a raceway section of a blast furnace, captured during an image-capture period that includes a period in which a furnace condition abnormality of the blast furnace occurred, is learned by a teacherless neural network; a second step in which, for each neuron constituting the teacherless neural network after learning, a correlation coefficient between an ignition value of the neuron and an index indicating furnace abnormality is calculated; and a third step in which the neurons to be used to detect furnace abnormalities are extracted as neurons for abnormality detection on the basis of the correlation coefficient.

Description

Blast furnace furnace condition learning method, furnace condition learning device, abnormality detection method, abnormality detection device and operation method

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, there has been proposed a method for detecting abnormal conditions of a blast furnace using a neural network, which learns using data related to whether the conditions of a known blast furnace are normal or abnormal. Specifically, Patent Document 1 describes a method for judging the state of a blast furnace tuyere using a neural network based on feature values extracted from images of the air path area and sensory inspection results by inspectors As a teaching material for learning, the image of the wind path area is captured by a camera device whose focus has been adjusted in advance. [Prior Art Literature] [Patent Literature]

Patent Document 1: Japanese Patent No. 6350159

(The problem that the invention intends to solve)

However, in the above method, if the causal relationship between the abnormality of the furnace condition to be detected and the measured data related to the furnace condition of the blast furnace is uncertain, or the measured If the time until the data on the condition is affected by the abnormality of the furnace condition to be detected is not clear, the learning data for detecting the abnormality of the furnace condition cannot be created, and the abnormality of the furnace condition cannot be detected with high accuracy.

The present invention has been made in view of the above-mentioned problems, and its object is to provide a furnace condition learning method and a furnace condition learning apparatus for a blast furnace, in which even if the causal relationship between the abnormality of the furnace condition to be detected and the measured data is not In the case of certainty, or when the time until the measured data is affected by the abnormality of the furnace condition to be detected is unclear, a neural network that can detect the abnormality of the furnace condition of the blast furnace with high accuracy can also be generated. 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 abnormality of the furnace condition of the blast furnace. (Technical means to solve problems)

The method for learning the furnace condition of the blast furnace of the present invention includes: a first step, which utilizes a non-teaching neural network to learn the image data of the blast furnace air path section, and the image data is included during the occurrence of abnormal furnace conditions of the blast furnace. Photographed during photography; the second step, which calculates the correlation coefficient between the excitation value of the neuron and the index representing the above-mentioned abnormality of the furnace condition for each neuron constituting the uninstructed neural network after learning; and the third step, which The neuron used for detecting the abnormality of the furnace condition based on the correlation coefficient is extracted as the abnormality detection neuron.

The above-mentioned first step may also include a step of learning a set of a plurality of image data of the time series of the wind path section by using a non-teaching neural network as one voxel data.

The above-mentioned first step may also include the following step, that is, a group of a plurality of image data of the above-mentioned wind path section captured by changing the focus of the photographing device is learned by using a non-teaching type neural network as a three-dimensional pixel data.

The apparatus for learning the furnace condition of the blast furnace according to the present invention includes means for learning image data of the blast furnace air path section by using a non-teaching type neural network, and the image data is captured during a period including the occurrence of abnormal furnace conditions of the blast furnace. Photographed during the period; means for calculating the correlation coefficient between the excitation value of the neuron and the index indicating the abnormal condition of the above-mentioned furnace for each neuron constituting the uninstructed neural network after learning; and extracting the above-mentioned correlation coefficient to detect the above-mentioned furnace The neuron used when the situation is abnormal is used as a means of abnormal detection of neurons.

The blast furnace abnormality detection method of the present invention includes the following steps: inputting the image data of the air path section photographed during the operation of the blast furnace into the uninstructed neural network learned by the blast furnace condition learning method of the present invention The steps of the network; and the step of detecting the abnormality of the furnace condition of the blast furnace according to the excitation value of the abnormality detection neuron.

The blast furnace abnormality detection device of the present invention includes: inputting the image data of the air path section captured during the operation of the blast furnace into the non-teaching type neural network learned by the blast furnace condition learning device of the present invention and a method for detecting abnormality of the furnace condition of the blast furnace based on the excitation value of the above-mentioned abnormality detection neuron.

The operating method of the blast furnace of the present invention includes the steps of operating the blast furnace while monitoring the abnormality of the furnace condition of the blast furnace by the abnormality detection method of the blast furnace of the present invention. (Compared to the efficacy of the prior art)

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

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

[Constitution of a blast furnace] First, with reference to FIGS. 1-3, the structure of the blast furnace to which the abnormality detection apparatus of a blast furnace which is one embodiment of this invention is applied is demonstrated.

FIG. 1 is a schematic diagram showing a configuration example of a blast furnace to which an abnormality detection apparatus for a blast furnace is applied, which is one embodiment of the present invention. As shown in FIG. 1 , a blower for blowing hot air from a hot blast stove (not shown) into the blast furnace 1 is connected to the inner side of the tuyere 2 of the blast furnace 1 to which the blast furnace abnormality detection device, which is one embodiment of the present invention, is applied. A pipe (blow pipe) 3 is provided, and a nozzle pipe 4 is provided to penetrate through the air supply pipe 3 . Fuels such as pulverized coal, oxygen, and city gas can be blown into the blast furnace 1 from the nozzle 4 . In the coke accumulation layer in front of the hot air supply direction of the tuyere 2, there is a combustion space called the air path area 5, and the coke combustion and gasification (reduction of iron ore, that is, the production of pig iron) is mainly in the air path area 5. conduct. Moreover, as shown in FIG. 2, the window 6 for furnace interior monitoring for an operator to monitor the situation in the blast furnace 1 is formed in the ventilation duct 3. As shown in FIG. Furthermore, in the vicinity of the window 6 for monitoring inside a furnace, the imaging device 7 which captures the image of the air path area|region 5 through the window 6 for monitoring inside a furnace is provided.

FIG. 3 is a schematic diagram showing an example of an image of the wind path region 5 captured by the photographing device 7 . As shown in FIG. 3 , in the image of the air path area 5 captured by the photographing device 7 , the nozzle 4 and the nozzle 4 and Outline of wind path zone 5. About 40 tuyere 2 are provided in the circumferential direction of the blast furnace 1 , and the images of the 40 air path areas 5 in the circumferential direction of the blast furnace 1 are taken one by one by the photographing device 7 . When the focal point of the photographing device 7 is adjusted, the position where the focal point is located inside the wind path area 5 changes. Therefore, when the focus is on the vicinity of the tuyere 2, the combustion state of the pulverized coal blown from the nozzle 4 can be easily observed, and when the focus is shifted from the tuyere 2 to the inside of the air path area 5, the coke can be easily observed. The internal state of the wind path area 5 such as the swirl state. In addition, when the time-series images of the wind path area 5 captured by the photographing device 7 are observed, changes in the combustion state of pulverized coal and changes in the swirl state of coke can be confirmed. In this way, the focal position of the photographing device 7 is adjusted, and the photographing is performed under a plurality of focal conditions from the tuyere 2 to the air path area 5, whereby the photographed time-series images and the like include the interior of the blast furnace 1. information about changes in the situation.

[Configuration of Abnormality Detection Device] Next, with reference to FIG. 4, the structure of the abnormality detection apparatus of a blast furnace which is one embodiment of this invention is demonstrated.

Fig. 4 is a block diagram showing the configuration of an abnormality detection apparatus for a blast furnace, which is one embodiment of the present invention. As shown in FIG. 4 , an abnormality detection device for a blast furnace, which is one embodiment of the present invention, includes an information processing device 11 , a display device 12 , and an operation condition adjustment device 13 .

The information processing device 11 is constituted by an information processing device such as a computer, and an arithmetic processing device such as a central processing unit (CPU, Central Processing Unit) inside the information processing device executes a computer program as a pre-learning processing unit 11a and an abnormality detection unit. The part 11b functions. The functions of the pre-learning processing unit 11a and the abnormality detecting unit 11b will be described later.

The display device 12 is constituted by 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 based on an operating input signal from an operator.

In the blast furnace abnormality detection device having such a configuration, the information processing device 11 executes the prior learning process and the abnormality detection process shown below, thereby generating a neural network capable of accurately detecting the abnormality of the furnace condition of the blast furnace 1, and The abnormality of the furnace condition of the blast furnace 1 is detected with high accuracy. Hereinafter, the operation of the information processing device 11 when the pre-learning process and the abnormality detection process are executed will be described.

[Pre-learning processing] First, referring to FIG. 5, the operation of the information processing apparatus 11 when the pre-learning process is executed will be described.

The abnormality of the furnace condition of the blast furnace 1, for example, indicates that the ventilation in the blast furnace 1 is deteriorated, and can be judged by using a furnace condition index such as a ventilation resistance index. If there is some kind of physical influence in the vicinity of the tuyere 2 corresponding to this abnormal furnace condition, the physical influence should be captured in the image of the air path area 5 captured by the photographing device 7 . However, since the internal phenomenon of the blast furnace 1 is not fully understood, it is not known whether the physical influence can be observed in the vicinity of the tuyere 2 . Also, even though it is known that there is a physical influence, it is not known to what extent the physical influence differs between the time detected by the furnace condition index and the time observed near the tuyere 2, so the furnace cannot be The condition index corresponds to the image data of wind path area 5. This means that in the case of applying machine learning, it is impossible to create teaching materials that can determine whether the judgment of normality or abnormality is correct.

Therefore, the information processing device 11 executes a pre-learning process of adjusting the parameters of each neuron constituting the neural network without creating teaching data. FIG. 5 is a flowchart showing the flow of the pre-learning process, which is one embodiment of the present invention. As shown in FIG. 5 , in the pre-learning process, which is an embodiment of the present invention, first, the pre-learning processing unit 11 a collects image data including the tuyere 2 that is continuous in time when the furnace condition is poor, which is captured by the photographing device 7 group (step S1). Next, the pre-learning processing unit 11a uses the image data set of the tuyere 2 collected in the processing of step S1 to perform unteaching learning on the neural network (step S2). Specifically, the pre-learning processing unit 11a takes the image data of the tuyere 2 as input, and reconstructs the output image data into the input image data set of the tuyere 2. In this way, through repeated calculation Optimizing the parameters of each neuron constituting the uninstructed neural network. Therefore, the excitation value of each neuron is adjusted in such a way that the value of a part of the image data set of the tuyere 2 used in the uninstructed learning becomes larger, but it is unknown to which image feature each neuron corresponds to. .

Therefore, the pre-learning processing unit 11a searches for neurons whose excitation values fluctuate in response to abnormal furnace conditions by calculating the correlation coefficient between the excitation values of neurons and the time-series data of the furnace condition index. Specifically, the deviation (time delay) T between the time when the abnormal furnace condition is detected by the furnace condition index and the time when the abnormal furnace condition is detected near the tuyere 2 is unknown. Therefore, first, the pre-learning processing unit 11a obtains the time delay Tmax at which the absolute value of the correlation coefficient becomes the largest while assigning the value of the time delay T to the excitation value of each neuron and the absolute value of the correlation coefficient at this time. Rmax (step S3). Next, the pre-learning processing unit 11a compares the absolute value Rmax of the correlation coefficient obtained between neurons for each neuron, and extracts the neuron for which the maximum value Rmax2 of the absolute value Rmax of the correlation coefficient is obtained (at this time, the absolute value Rmax of the correlation coefficient is extracted). The time delay is set to Tmax2) (step S4). Finally, the pre-learning processing unit 11a compares the furnace condition index with the excitation value of the extracted neuron to determine the threshold value S of the excitation value of the neuron that can be extracted while the furnace condition is abnormal (step S5). Thereby, a series of pre-learning processing is completed.

In addition, the maximum value Rmax2 of the correlation coefficient is the value calculated from the neuron having the highest correlation with the furnace condition index among the neurons after the uninstructed learning 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 judged that the image data set of the tuyere 2 reflects the influence of abnormal furnace conditions. In addition, under the condition that the excitation value of the neuron changes earlier than the furnace condition index, and the time delay Tmax2 is obtained, the excitation value of the neuron can be used to predict the abnormality of the furnace condition in advance. In other words, if the time delay Tmax2 is greater than zero when the correlation coefficient is calculated by staggering the firing value of the neuron with respect to the furnace condition index by the time Tmax2, the firing value of the neuron can be used to predict the furnace condition in advance. abnormal. As described above, in order to utilize the firing value of neurons in abnormal detection of blast furnace 1, the firing value of neurons must be changed prior to the furnace condition index. Therefore, when searching for the maximum value Rmax2 of the correlation coefficient, it is preferable to perform the search under the condition that the time delay Tmax2 is greater than zero.

[Anomaly detection processing] Next, referring to FIG. 6 , a flow of abnormality detection processing according to an embodiment of the present invention will be described.

FIG. 6 is a flowchart showing a flow of abnormality detection processing according to an embodiment of the present invention. As shown in FIG. 6 , in the abnormality detection process, which is one embodiment of the present invention, first, the imaging device 7 captures the image data of the tuyere 2 (step S11 ). Next, the abnormality detection unit 11b inputs the image data of the tuyere 2 captured in the process of step S11 into the neural network that has been learned by the pre-learning process, thereby calculating the excitation of each neuron constituting the neural network value (step S12). Next, when the excitation value of the neuron is greater than or equal to the threshold value S determined in advance, the abnormality detection unit 11b determines that the abnormality of the furnace condition corresponding to the neuron occurs after the time Tmax, and notifies the operator of the abnormality. The warning information of the content is output to the display device 12 (step S13). The operator operates the operating condition adjusting device 13 according to the warning information displayed on the display device 12 , thereby adjusting the operating conditions of the blast furnace 1 . According to this process, the operation state of the blast furnace 1 can be quickly adjusted in response to the warning information, so that the abnormality of 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 actual tuyere image data and abnormal furnace conditions will be described. For a blast furnace with 40 tuyere in the circumferential direction, a photographic device was installed and video data were collected. During the 11-day period, including the abnormal furnace condition, one image was extracted every hour from each tuyere to obtain image data and set as an image data group. The image data is set as RGB (Red Green Blue, Red Green Blue) color image of 320×240 pixels. In addition, ventilation deterioration was selected as the furnace condition abnormality, and the ventilation resistance index was used as the furnace condition index. In addition, the neural network is a non-teaching type neural network using the 12-layer structure described in the literature (Le et al., Building High-level Features Using Large Scale Unsupervised Learning, ICML 2012). Furthermore, the neural network is a locally coupled neural network, but it can also be an overall coupled neural network, CNN (Convolutional Neural Network, convolutional neural network), GAN (Generative Adversarial Network, generative adversarial network) A neural network for unsupervised learning using video data.

According to the flow chart shown in FIG. 5 , after the input of the image data set, the uninstructed learning of the neural network is implemented. Here, the calculation method of the correlation coefficient will be described. First, the time-series images of each tuyere are input to the neural network after learning without instruction, and the excitation value of each neuron is obtained for each tuyere. Second, normalize the firing value of each neuron. As normalization methods, min-max normalization using the maximum value and minimum value and z-score normalization using the mean value and standard deviation are known, but in this embodiment, to use the former. However, the latter can also be used. The data of the normalized excitation value of each neuron are averaged among the tuyere to obtain the average excitation value of all the tuyere. Furthermore, in this embodiment, the average excitation value of all the tuyere is obtained by averaging the output of the excitation value of the neuron among the tuyere, but it is also possible to juxtapose the images of all 40 cameras into one image. The image data may be input later as the excitation value of the neurons of the neural network after learning.

The average excitation value and furnace condition index of all the tuyere are distributed under the condition that the time delay T is 0 < time delay T < 8 hours, and the absolute value of the correlation coefficient becomes the maximum time delay Tmax and the correlation coefficient at this time. The absolute value of Rmax. Here, 8 hours is the average cycle time until the raw material put into the blast furnace becomes molten iron, and it is considered that there is no physical causal relationship with respect to a longer time. Then, the absolute value Rmax of the correlation coefficient is compared between the neurons, and the neuron having the maximum value Rmax2 of the absolute value Rmax of the correlation coefficient is obtained and extracted. At this time, the maximum value Rmax2 of the absolute value Rmax of the correlation coefficient is 0.41, and the time delay Tmax2 at this time is 1 hour. In general, when the correlation coefficient is about 0.4, although the correlation is not too large, it is considered that not all the factors of the ventilation deterioration are observed in the tuyere, and thus it is judged to be a sufficient correlation coefficient.

FIG. 7 shows a graph in which the ventilation resistance index and the time change of the average neuron firing value of all the tuyere are recorded together. Neuron firing values are depicted with a 1-hour stagger, taking into account the time delay Tmax2. As shown in Figure 7, when the ventilation resistance index is high, the furnace condition is judged to be poor, and when the ventilation resistance index exceeds 1.1, the degree of abnormal furnace condition is particularly high. In the second half of the date of 1/17, the ventilation resistance index rose sharply, and it was known that the ventilation became particularly poor. Although the figure shows that the neuron excitation value rises 1 hour earlier than the ventilation resistance index during the exploration process, but after qualitatively comparing the coordinate graph shown in Figure 7, it can be seen that it rises to around 1.1 above the arrows R1 and R2. , the neuron firing values were higher from the first few hours. Therefore, if the threshold value S of the neuron excitation value is determined to be 0.26, for example, the blast furnace operation in abnormal furnace conditions can be suppressed in advance, such as reducing the air supply amount before occurrence of abnormal furnace conditions. In this way, even in the case where the causal relationship between the ventilation of the detection object and the captured image is uncertain, the neurons can learn the feature quantity of the image through uninstructed learning in advance, and by exploring and matching in the next step. Neurons with higher ventilation correlations obtain metrics that can be used for anomaly detection. In addition, by searching for the time delay T, it is possible to clarify the timing until the ventilation is affected by the captured image.

Furthermore, one image usually captures the situation in the furnace at a certain moment, so the furnace information obtained from the image is the spatial hue change that can be observed in the air path area. On the other hand, abnormal furnace conditions such as ventilation deterioration in the furnace are information of temporal changes. Therefore, the image data is processed by integrating a plurality of images continuously shot in time series into one voxel data, which can introduce the information of the change of color tone with time, so the detection performance can be improved. For example, in the image data in the above study example, 10 images (10 seconds) of image data obtained by shooting once per second are integrated into the volume pixel data, as long as the above-mentioned image data are collected at 40 air outlets every 1 hour, etc. Can.

Also, when the focus of the photographing device is fixed, only information of a certain depth of the air path region having a depth in the furnace direction can be obtained. Therefore, the image data is processed into one voxel data by integrating a plurality of images captured at a plurality of focal points, and information in the depth direction can be introduced, so that the detection performance can be improved. For example, the image data in the above learning example is set as the voxel data of 10 images captured after adjusting the focus to 10 stages, respectively, and the stereoscopic data is collected at 40 air outlets every 1 hour. Pixel data, etc. In addition, both the time-series image and the image after changing the focus may be used. In this case, for example, the focus is divided into 10 stages, and 10 images (10 seconds) are captured by shooting once per second, and all the image data are integrated into voxel data, which is used as the image in the above learning example Data, and voxel data can be collected every 1 hour at the 40 air outlets.

As mentioned above, although the embodiment to which the invention made by the inventors of the present invention is applied has been described, the present invention is not limited to the descriptions and drawings which constitute a part of the disclosure of the invention by this embodiment. That is, according to the present embodiment, all other embodiments, examples, operating techniques, etc., are all included in the scope of the present invention for those skilled in the art. (Industrial Availability)

According to the present invention, it is possible to provide a furnace condition learning method and a furnace condition learning device for a blast furnace, in which even if the causal relationship between the abnormal furnace condition to be detected and the measured data is not clear, or the measured In the case where the time when the output data is affected by the abnormal furnace condition to be detected is not clear, the neural network for detecting the abnormal furnace condition of the blast furnace with high accuracy can also provide the furnace condition learning method and the furnace condition learning device of the blast furnace. . Furthermore, according to the present invention, it is possible to provide a blast furnace abnormality detection method, an abnormality detection device, and an operation method capable of accurately detecting abnormality of the furnace condition of the blast furnace.

1: blast furnace 2: Air outlet 2a: Small air outlet 3: Air supply pipe (blow pipe) 4: Nozzle 5: Wind path area 6: Window for monitoring in the furnace 7: Photographic installation 11: Information processing device 11a: Prior Learning Processing Section 11b: Anomaly Detection Section 12: Display device 13: Operating condition adjustment device

FIG. 1 is a schematic diagram showing a configuration example of a blast furnace to which an abnormality detection device for a blast furnace is applied, which is one embodiment of the present invention. FIG. 2 is a schematic diagram showing a configuration example of a blast furnace to which an abnormality detection apparatus for a blast furnace is applied, which is an embodiment of the present invention. FIG. 3 is a schematic diagram showing an example of an image of a wind path area captured by a photographing device. Fig. 4 is a block diagram showing the configuration of an abnormality detection apparatus for a blast furnace, which is one embodiment of the present invention. FIG. 5 is a flowchart showing the flow of the pre-learning process, which is one embodiment of the present invention. FIG. 6 is a flowchart showing a flow of abnormality detection processing according to an embodiment of the present invention. FIG. 7 is a graph showing an example of the temporal change of the ventilation resistance index and the average neuron firing value of all the tuyere.

Claims (7)

A furnace condition learning method of a blast furnace, comprising: Step 1, which utilizes a non-teaching neural network to learn the image data of the blast furnace air path section, the image data being taken during the photographing period including the occurrence period of the abnormal furnace condition of the blast furnace; The second step is to calculate the correlation coefficient between the firing value of the neuron and the index indicating the abnormal furnace condition for each neuron constituting the learned uninstructed neural network; and In the third step, the neuron used for detecting the abnormality of the furnace condition based on the correlation coefficient is extracted as the neuron for abnormality detection. The method for learning the furnace condition of a blast furnace according to claim 1, wherein the first step includes learning a set of a plurality of image data in the time series of the air path section by using a non-teaching neural network as 1 voxel data step. The method for learning the furnace condition of a blast furnace according to claim 1 or 2, wherein the first step comprises: using a non-teaching neural network for a set of a plurality of image data of the air path section captured by changing the focus of the photographing device The way is learned as a step of 1 voxel data. A furnace condition learning device for a blast furnace, comprising: A means of using an uninstructed neural network to learn the image data of the blast furnace air path section, the image data being taken during the photographing period including the period during which the abnormal condition of the blast furnace occurred; means for calculating the correlation coefficient between the firing value of the neuron and the index representing the above-mentioned abnormality of the furnace condition for each neuron constituting the uninstructed neural network after learning; and The neuron used when the abnormality of the furnace condition is detected based on the correlation coefficient is extracted as a means of abnormality detection neuron. An abnormality detection method for a blast furnace, comprising the following steps: The step of inputting the image data of the air path section taken during the operation of the blast furnace into the uninstructed neural network learned by the furnace condition learning method of the blast furnace in any one of claim 1 to 3; and The step of detecting the abnormality of the furnace condition of the blast furnace according to the excitation value of the abnormality detection neuron. An abnormality detection device for a blast furnace, comprising: The means of inputting the image data of the air path section taken during the operation of the blast furnace into the non-teaching type neural network learned by the furnace condition learning device of the blast furnace of claim 4; and A means for detecting abnormality in the furnace condition of the blast furnace based on the excitation value of the abnormality detection neuron. An operation method of a blast furnace, comprising the steps of operating the blast furnace while monitoring the abnormality of the furnace condition of the blast furnace using the blast furnace anomaly detection method of claim 5.

Images