A method and system for online intelligent sensing of the state of molten iron flow in a blast furnace

Abstract

This invention discloses an online intelligent sensing method and system for the state of molten iron flow in a blast furnace. By acquiring video frames of molten iron flow during the tapping process, a ResNeXt network based on deep convolution is constructed to obtain an abnormal state recognition model for the molten iron flow. Abnormal states of the molten iron flow are detected. Based on the recognition results, a quasi-abnormal time-series dataset is constructed. Dynamic features of the flow from the molten iron flow video frames are extracted, and high-dimensional features from the quasi-abnormal time-series dataset are extracted using a Conv_Lstm network. Simultaneously, a FNN network is used to extract high-dimensional features of the dynamic flow features. The high-dimensional features extracted by the two networks are then concatenated and fused to obtain an online intelligent sensing model for the state of the molten iron flow. This model enables the detection of the molten iron flow state and solves the technical problem of low accuracy in online recognition of the molten iron flow state during the tapping process in existing blast furnaces, achieving online intelligent sensing and accurate recognition of the molten iron flow state.

Description

TECHNICAL FIELD

[0001] The present application mainly relates to the technical field of blast furnace smelting, in particular to a method and system for online intelligent perception of molten iron flow state in a blast furnace. BACKGROUND

[0002] The blast furnace tapping process, as the "last link" of the blast furnace ironmaking, is the most important regulation and control link in the lower part of the blast furnace. The molten iron flow state in the tapping process is the most intuitive and important representation of the tapping state. Therefore, online intelligent perception of the molten iron flow state in the tapping process is crucial to improve the digital and intelligent level of the tapping operation. The pressure fluctuation of the furnace hearth, the dynamic changes of the slag-iron flowability and liquid level, the damage of the iron notch mud sleeve and mud bag, and the abnormal erosion and blockage of the iron notch, can easily lead to abnormal tapping states such as "spattering", "stuck coke", and "large flow". When the iron notch spatters, the high-pressure gas in the furnace escapes from the iron notch, which reduces the pressure in the furnace bosh and easily leads to material collapse accidents, accelerates the erosion of the iron notch, and affects the safety and service life of the iron notch. When the iron notch is stuck with coke, the discharge of slag and iron is not smooth, which significantly reduces the tapping efficiency, and further leads to abnormal increase of the furnace hearth liquid level, difficulty in downward movement of the furnace charge, and sharp increase of the smelting energy consumption, which is not conducive to safe and efficient production. When the molten iron runs in a large flow, the instantaneous flow of slag and iron is too large, which exceeds the maximum safe flow of the main iron channel and skimmer, and will accelerate the erosion of the iron channel and skimmer. The molten iron flows into the slag channel and enters the slag pot, which may cause serious accidents such as explosion. Therefore, online intelligent perception of the molten iron flow state in the tapping process is of great significance to ensure safe, efficient and stable tapping, and to promote the stable operation of the blast furnace as a whole, energy saving and carbon reduction, and improvement of the quality of molten iron and production efficiency.

[0003] Currently, the intermittent observation of the molten iron flow state in the tapping process is performed by workers on a shift basis. Due to the high temperature, high dust, strong vibration and strong radiation of the molten iron in the tapping field, direct observation of the molten iron flow state is highly dangerous and has a great damage to the vision. The current tapping site cannot continuously and intelligently perceive the molten iron flow state information, and the operation relies heavily on the subjective experience of workers. The regulation and control measures are extensive, lagging and blind, which seriously restricts the safe and stable operation, energy saving and consumption reduction, and quality improvement and efficiency increase of the blast furnace smelting process. Therefore, online intelligent perception of the molten iron flow state in the tapping process will greatly reduce the labor intensity of workers in the tapping site, and revolutionize the production mode of the existing tapping process with high risk coefficient, low production efficiency and lagging operation decision-making. It lays a foundation for iron notch maintenance, real-time abnormal disposal, automatic and intelligent opening / closing of the iron notch in the tapping process, and promotes the transformation and upgrading of the tapping process to "intelligent tapping".

[0004] The object of the application is the state of the molten iron stream of the blast furnace. There are extremely high temperature, strong radiation, dynamic dust and other harsh environmental disturbances in the tapping site, and the molten iron stream has high speed and strong light, which greatly increases the difficulty of identifying the state of the molten iron stream. The prior art mainly realizes the identification of the state of the molten iron stream by directly observing the molten iron stream by artificial observation. Artificial direct observation has great damage to vision, cannot observe for a long time, and artificial observation cannot observe the details of the texture and morphological characteristics of the molten iron stream, and the observation accuracy and timeliness are low. On the other hand, the area of the molten iron stream in the tapping process is detected by video information to realize the monitoring of the state of the tapping hole. However, this method can only realize the state monitoring of the tapping hole, and cannot realize the real-time monitoring of the stream state (especially the abnormal stream state).

[0005] The invention patent with the publication number CN113122669B discloses a blast furnace tapping hole state monitoring method and system, which adopts the image of the tapping hole of the blast furnace, and performs morphological operations such as image interception, morphological opening operation, Gaussian blur, color clustering, gray scale conversion and threshold segmentation on the image to obtain the image of the molten iron stream of the blast furnace tapping hole. The size and depth of the tapping hole are monitored by calculating the corresponding molten iron area and time parameters in the feature image. Due to the strong light and dust interference in the actual site, the accuracy of the molten iron area segmentation needs to be verified, and only the area information of the molten iron area cannot realize the online intelligent perception of the stream state of the molten iron stream in the tapping process. SUMMARY

[0006] The application provides a blast furnace molten iron stream state online intelligent perception method and system, which solves the technical problem of low online identification accuracy of the stream state of the molten iron stream in the existing blast furnace tapping process.

[0007] To solve the above technical problems, the application provides a blast furnace molten iron stream state online intelligent perception method, which comprises the following steps:

[0008] Collecting the video frames of the molten iron stream in the tapping process, and constructing the molten iron stream state data set of different stream state categories.

[0009] Constructing a ResNeXt network based on deep convolution to obtain a molten iron stream stream abnormal state recognition model, and detecting the molten iron stream stream abnormal state according to the molten iron stream stream abnormal state recognition model.

[0010] According to the recognition result of the molten iron stream stream abnormal state, a quasi-abnormal time series data set is constructed.

[0011] Extracting the stream dynamic characteristics of the molten iron stream video frames.

[0012] The Conv_Lstm network is used to extract high-dimensional features of the quasi- abnormal time series data set, the FNN network is used to extract high-dimensional features of the stream dynamic characteristics, and the high-dimensional features extracted by the two networks are spliced and fused to obtain an online intelligent perception model of the stream state, and the stream state is detected according to the online intelligent perception model of the stream state.

[0013] Further, the video frames of the stream of molten iron in the tapping process include:

[0014] The tapping state of the tapping hole is identified, and the specific identification formula is:

[0015]

[0016] Wherein, T is the tapping state of the tapping hole, 1 represents that the tapping is in progress, 0 represents that the tapping is not in progress, I represents the video frame of the stream of molten iron, max(I) represents the maximum value of the pixels in the video frame of the stream of molten iron, OTSU(I) represents the image threshold value of the OTSU algorithm, sum(I>OTSU(I)) represents the number of pixels in the image that exceeds the threshold value, ω, ξ, τ respectively represent the first threshold value parameter, the second threshold value parameter and the third threshold value parameter of the self-defined tapping state detection.

[0017] According to the identification result of the tapping state of the tapping hole, the video frames of the stream of molten iron in the tapping process are collected.

[0018] Further, after collecting the video frames of the stream of molten iron in the tapping process, it further includes:

[0019] The video frames of the stream of molten iron are angle transformed and equidimensional scaled.

[0020] Further, the skewness loss function of the stream of molten iron stream abnormal state recognition model is specifically:

[0021]

[0022] Wherein, L represents the loss function of the sample in the training process, p A and p N are the probabilities of the stream of molten iron stream abnormal state recognition model being predicted as an abnormal stream state and a normal stream state, n A is the number of samples of the stream of molten iron abnormal state, n N is the number of samples of the normal stream state of the stream of molten iron, θ respectively are the first modulation coefficient and the second modulation coefficient of the loss function, y is the label value of the current video frame of the stream of molten iron, and when y=1 represents the stream of molten iron stream abnormal state, y=0 represents the stream of molten iron stream normal state.

[0023] Further, according to the identification result of the stream of molten iron stream abnormal state, the quasi- abnormal time series data set is constructed, which includes:

[0024] According to the identification result of the abnormal state of the molten iron flow stream, a quasi- abnormal sample is obtained, and the quasi- abnormal sample includes a true abnormal sample identified as an abnormal sample and a normal sample incorrectly identified as an abnormal sample.

[0025] According to the time sequence corresponding to the sample in the quasi- abnormal sample, a quasi- abnormal time sequence data set is obtained, and the time sequence corresponding to a single sample is specifically a preset number of image frames contained in the current video frame in time reverse order.

[0026] Further, the stream dynamic feature of the molten iron flow video frame includes:

[0027] Extracting the stream trajectory of the molten iron flow video frame.

[0028] Obtaining the stream boundary fluctuation rate and / or the convex feature of the stream trajectory as the stream dynamic feature of the molten iron flow video frame.

[0029] Further, the specific formula of the stream boundary fluctuation rate of the stream trajectory is:

[0030]

[0031] Wherein, V represents the stream boundary fluctuation rate, λ represents the stream aspect ratio coefficient, d t represents the actual size of the steel nozzle at the taphole, l s represents the actual size of a pixel point of the current video frame corresponding to the photographed object, L t (i) represents the corresponding longitudinal coordinate of the upper boundary of the molten iron flow stream at the horizontal coordinate i of the molten iron flow image corresponding to the molten iron flow video frame, y t (i) represents the corresponding value of the upper boundary parabola equation of the molten iron flow at the horizontal coordinate i of the molten iron flow image, L b (i) represents the corresponding longitudinal coordinate of the lower boundary of the molten iron flow stream at the horizontal coordinate i of the molten iron flow image, y b (i) represents the corresponding value of the lower boundary parabola equation of the molten iron flow at the horizontal coordinate i of the molten iron flow image.

[0032] Further, the convex feature of the stream trajectory includes:

[0033] A stream convex boundary pixel horizontal coordinate set B = {b1, b2,..., b r} is constructed, wherein:

[0034]

[0035] Wherein, j represents the serial number of the stream convex boundary pixel horizontal coordinate set, r represents the number of boundary pixels contained in the convex boundary pixel horizontal coordinate set, L t (b j) represents the ordinate corresponding to the upper boundary of the stream of liquid iron at the abscissa b j of the stream of liquid iron, L b (b j ) represents the ordinate corresponding to the lower boundary of the stream of liquid iron at the abscissa b j of the stream of liquid iron, y t (b j ) represents the value corresponding to the parabolic equation of the upper boundary of the stream of liquid iron at the abscissa b j of the stream of liquid iron, y b (b j ) represents the value corresponding to the parabolic equation of the lower boundary of the stream of liquid iron at the abscissa b j of the stream of liquid iron.

[0036] The number of sets of protruding pixels of the boundary of the stream of liquid iron in the video frame of the stream of liquid iron is recorded, and the aspect ratio feature of the protruding boundary is calculated, and the specific calculation formula is:

[0037]

[0038] Where R is the aspect ratio of the protruding boundary feature, l is the distance between b r , and D l (b j , l) is the distance from the boundary pixel point corresponding to b j to the straight line l, D p (b1, b r ) is the Euclidean distance of the two boundary pixel points corresponding to b r .

[0039] The present application provides a kind of blast furnace stream of liquid iron stream state online intelligent perception system, comprising:

[0040] Memory, processor and computer program stored in memory and can be run on processor, when processor executes computer program, the steps of the method for realizing the blast furnace stream of liquid iron stream state online intelligent perception method provided by the present application.

[0041] The application provides a blast furnace molten iron flow state online intelligent sensing method and system, which acquires molten iron flow video frames in the tapping process, constructs a ResNeXt network based on deep convolution to obtain a molten iron flow stream abnormal state recognition model, detects the molten iron flow stream abnormal state, constructs a quasi-abnormal time series data set according to the recognition result of the molten iron flow stream abnormal state, extracts stream dynamic features of the molten iron flow video frames and high-dimensional features of the quasi-abnormal time series data set by using a Conv_Lstm network, simultaneously extracts high-dimensional features of the stream dynamic features by using an FNN network, splices and fuses the high-dimensional features extracted by the two networks to obtain a molten iron flow stream state online intelligent sensing model, and detects the molten iron flow stream state according to the molten iron flow stream state online intelligent sensing model, thereby solving the technical problem of low online recognition precision of the molten iron flow stream state in the existing blast furnace tapping process, realizing online intelligent sensing and accurate recognition of the molten iron flow stream state, and having the advantages of high precision, strong stability, good real-time performance, low investment, convenient installation, long service life and the like, and filling the gap in the monitoring of the molten iron flow stream state in the current blast furnace tapping field.

[0042] The purposes of the application include:

[0043] (1) The application acquires real-time blast furnace tapping hole molten iron flow images, and proposes a two-stage skewness prediction and feature fusion molten iron flow stream state recognition strategy, so as to realize online intelligent sensing of the molten iron flow stream state.

[0044] (2) In view of the deficiencies in the field, the purpose of the application is to design a blast furnace tapping process molten iron flow stream state intelligent sensing method and system, which acquires tapping process molten iron flow high-speed video frames by using a high-speed CMOS sensor. By constructing a data set of different stream state categories, high-precision detection of the molten iron flow stream abnormal state is realized by combining a skewness loss function; meanwhile, a dynamic construction of a quasi-abnormal data set and a stream dynamic / steady feature extraction method are proposed, and the time series image sequence and the stream features are fused, so as to finally realize online intelligent sensing of the molten iron flow stream state. The method and system have the advantages of high precision, strong stability, good real-time performance, low investment, convenient installation, long service life and the like, and fill the gap in the monitoring of the molten iron flow stream state in the current blast furnace tapping field.

[0045] (3) The application aims to provide a non-contact molten iron flow state online real-time intelligent sensing method and system which acquires real-time molten iron flow video frames by using a high-speed CMOS sensor.

[0046] (4) The application aims to provide a molten iron flow stream state non-uniform distribution situation, molten iron flow stream shielding state detection and molten iron flow stream abnormal state efficient detection method.

[0047] (5) The purpose of the present application is to provide a molten iron flow stream quasi-abnormal state data set construction method and a molten iron flow stream state two-stage high-precision identification method.

[0048] (6) The purpose of the present application is to provide a molten iron flow stream abnormal state identification and abnormal state early warning method, which overturns the model of manual observation of molten iron flow state on site, and provides real-time data support for timely disposal of abnormal states on site.

[0049] The beneficial effects of the present application specifically include:

[0050] (1) The present application uses a high-speed CMOS sensor to collect real-time molten iron flow video at the blast furnace tapping hole, and proposes a two-stage intelligent perception method for the molten iron flow stream state. In view of the uneven characteristics of normal samples and abnormal samples of the molten iron flow stream state, a skew instantaneous function is proposed in the first stage, and a deep CNN network is constructed to realize efficient detection of the abnormal state of the molten iron flow. At the same time, in the second stage, a molten iron flow stream quasi-abnormal state time series data set is constructed, and a molten iron flow stream dynamic feature acquisition method is proposed, which realizes efficient fusion of the quasi-abnormal data and dynamic features through the Conv_Lstm network and the FNN network, and finally realizes online intelligent perception of the molten iron flow stream state category at the blast furnace tapping hole.

[0051] (2) The present application proposes a molten iron flow stream consistency correction and equal-scale scaling strategy to solve the problems of image distortion, inconsistent scales and low data quality caused by uncontrollable installation pose, distance and iron hole exit angle of industrial field imaging equipment, and realizes molten iron flow pose distortion correction, molten iron flow equal-scale scaling, flow stream exit angle detection and correction.

[0052] (3) The present application constructs an abnormal skew loss function in view of the extremely uneven characteristics of the number of normal state and abnormal state samples of the molten iron flow stream in the actual tapping process, effectively solves the problem of prediction result biasing to the normal state based on model inference confidence and sample size constraints, and realizes efficient detection of the abnormal state of the molten iron flow stream.

[0053] (4) The present application proposes flow stream trajectory morphological features, flow stream boundary fluctuation features, etc. to effectively obtain typical feature information of the molten iron flow stream, and finally realizes online intelligent perception of the molten iron flow stream state by inputting the features into FNN for deep feature acquisition and fusion with high-dimensional time series features of Conv_Lstm. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 The present application is a blast furnace tapping process molten iron flow stream state online intelligent perception method flow chart for the second embodiment of the present application.

[0055] Figure 2 A hardware schematic diagram for high-speed acquisition of molten iron flow of the second embodiment of the present application;

[0056] Figure 3 A camera mounting position schematic diagram of the second embodiment of the present application;

[0057] Figure 4 A structural block diagram of the online intelligent perception system for the flow state of the molten iron flow of the blast furnace of the present application.

[0058] Reference signs:

[0059] 1, molten iron flow; 2, high-speed imaging device; 3, support; 4, 10-gigabit Ethernet; 5, high-performance computer; 10, memory; 20, processor. DETAILED DESCRIPTION

[0060] In order to facilitate the understanding of the present application, the present application will be described more fully, specifically and concretely below with reference to the accompanying drawings and preferred embodiments, but the scope of protection of the present application is not limited to the following specific embodiments.

[0061] The embodiments of the present application will be described in detail below with reference to the accompanying drawings, but the present application can be implemented in various different ways limited and covered by the claims.

[0062] Embodiment one

[0063] The online intelligent perception method for the flow state of the molten iron flow of the blast furnace provided by the first embodiment of the present application comprises:

[0064] Step S101, collect the video frames of the molten iron flow in the tapping process, and construct the molten iron flow state data set of different flow state categories.

[0065] Step S102, construct a ResNeXt network based on deep convolution to obtain an abnormal state recognition model of the molten iron flow, and detect the abnormal state of the molten iron flow according to the abnormal state recognition model of the molten iron flow.

[0066] Step S103, construct a quasi-abnormal time series data set according to the recognition result of the abnormal state of the molten iron flow.

[0067] Step S104, extract the flow dynamic features of the video frames of the molten iron flow.

[0068] Step S105, use a Conv_Lstm network to extract high-dimensional features of the quasi-abnormal time series data set, use a FNN network to extract high-dimensional features of the flow dynamic features, and splice and fuse the high-dimensional features extracted by the two networks to obtain an online intelligent perception model for the flow state of the molten iron flow, and detect the flow state of the molten iron flow according to the online intelligent perception model for the flow state of the molten iron flow.

[0069] The online intelligent sensing method for the state of molten iron flow in a blast furnace provided in this invention collects video frames of the molten iron flow during the tapping process, constructs a ResNeXt network based on deep convolution to obtain an abnormal state recognition model for the molten iron flow, detects abnormal states of the molten iron flow, constructs a quasi-abnormal time-series dataset based on the recognition results, extracts the dynamic features of the molten iron flow from the video frames, and extracts high-dimensional features of the quasi-abnormal time-series dataset using a Conv_Lstm network. Simultaneously, an FNN network is used to extract high-dimensional features of the dynamic features of the molten iron flow. The system identifies and merges the high-dimensional features extracted from two networks to obtain an online intelligent sensing model for the state of molten iron flow. This model is then used to detect the state of the molten iron flow, solving the technical problem of low accuracy in online identification of the molten iron flow state during blast furnace tapping. It achieves online intelligent sensing and accurate identification of the molten iron flow state, offering significant advantages such as high accuracy, strong stability, good real-time performance, low investment, convenient installation, and long service life. This fills the current gap in molten iron flow state monitoring at blast furnace tapping sites.

[0070] Example 2

[0071] The present invention first establishes a dataset of the state of the molten iron flow stream during the tapping process, and then constructs a two-stage identification network model for the state of the molten iron flow stream to achieve online intelligent perception of the state of the molten iron flow stream during the tapping process.

[0072] like Figure 1 As shown, Figure 1 This is a flowchart of an online intelligent sensing method for the state of molten iron flow during the blast furnace tapping process, according to an embodiment of the present invention. Figure 1 The method for online intelligent sensing of molten iron flow status during blast furnace tapping includes the following steps: Step U1, high-speed acquisition of molten iron flow; Step U2, identification of molten iron flow region occlusion status; Step U3, image preprocessing and dataset construction; Step U4, identification of abnormal flow state skewness; Step U5, construction of quasi-abnormal dataset; Step U6, extraction of dynamic and steady-state features of the flow; Step U7, fusion of temporal state and flow feature; Step U8, data display and early warning.

[0073] The specific process is as follows: First, clear video frames of molten iron flow are acquired through high-speed image acquisition at the tapping site. Simultaneously, to avoid misjudgments of the molten iron flow state due to on-site dust and personnel operation, occlusion status identification is used to initially screen for molten iron flow obstruction. Unobstructed molten iron flow video frames proceed to the next identification stage. Next, image preprocessing and dataset construction are used to correct the relative position of the taphole, perform iso-scale scaling and image transformation, and combine this with manual state annotation to achieve dataset label matching. For video frames of molten iron flow state with non-equilibrium distribution, a deep residual convolutional network model (ResNext) is constructed through abnormal flow state skewness identification. To avoid the model's inference results being biased towards normal operating conditions due to imbalanced data samples, a skewness-constrained loss function is constructed to achieve efficient detection of quasi-abnormal states of the molten iron flow. For the detection... The quasi-anomaly image frames of the molten iron flow are used to construct a quasi-anomaly time-series dataset for the iron tapping process. To further improve the category recognition accuracy of the quasi-anomaly samples, effective state features of the molten iron flow are extracted through dynamic and steady-state characteristics of the flow. In the time-series state and flow feature fusion step, a Convolutional Long Short-Term Memory Network (Conv_Lstm) and a Feedforward Neural Network (FNN) are constructed for the constructed quasi-anomaly time-series dynamic dataset of the iron tapping process and the extracted effective state features of the molten iron flow, respectively. The high-level features of the two models are then weighted and concatenated to achieve real-time intelligent perception of the molten iron flow state. Finally, the recognition results and warning information of the model are further displayed on the software interface through the data display and warning steps, thereby realizing online intelligent perception of the molten iron flow state during the iron tapping process using the trained two-stage model. The specific steps are explained in detail below:

[0074] Step U1: High-speed collection of molten iron flow.

[0075] The hardware schematic diagram for high-speed molten iron flow acquisition in this embodiment is as follows: Figure 2 As shown, the system mainly includes a high-speed imaging device 2, a support 3, a 10 Gigabit Ethernet port 4, and a high-performance computer 5. Specifically, its front-end image acquisition hardware mainly includes a high-speed CMOS photosensitive device for capturing the molten iron flow stream 1, and a strong light suppression filter to prevent damage to the CMOS device and overexposure caused by the strong light from the molten iron flow. Simultaneously, the sensor's electrical signals are converted into optical signals via a photoelectric converter for long-distance transmission. Combined with a high-speed image acquisition card, the images are decoded to achieve high-speed image acquisition. The image acquisition equipment is equipped with a cooling and protection device, including a metal outer shell, compressed gas cooling purging, and a heat insulation layer, to ensure long-term stable operation of the equipment in high-temperature and dynamic dust environments at industrial sites.

[0076] Step U2: Identification of the occlusion status of the molten iron flow area.

[0077] In the actual tapping process, different tapholes of a blast furnace are periodically alternately tapped, and for a single taphole, there are two states of tapping and non-tapping. And in the tapping process, due to the operation of the site operator near the taphole and the random dust interference of the tapping field, the molten iron flow stream presents the possibility of being blocked. The video frames of the blocked molten iron flow need to be identified and removed to increase the robustness of the subsequent molten iron flow state identification model.

[0078] 1. Taphole tapping state identification

[0079] The tapping market of a single taphole is usually about 1-3 hours according to the volume and production load difference of the blast furnace, and the molten iron flow state identification in the tapping process is for the taphole of the blast furnace being tapped, so it is necessary to first automatically identify the tapping state of the tapping taphole:

[0080]

[0081] In the formula, T is the tapping state of the taphole, 1 represents tapping, 0 represents non-tapping, I represents the video frame of the molten iron flow, max(I) is the maximum value of the pixels in the molten iron flow video frame, OTSU(I) represents the image threshold value of the OTSU algorithm, sum(I>OTSU(I)) represents the number of pixels in the image that exceeds the threshold value. ω, ξ, τ respectively represent the first threshold parameter, the second threshold parameter and the third threshold parameter of the self-defined tapping state detection, which are related to the front-end hardware devices such as the camera installation pose, the type of CMOS photosensitive element, the lens focal length, the aperture and the like of the tapping field.

[0082] For the video frame of the molten iron flow in the tapping process, the molten iron flow presents a free jet state after being ejected from the taphole, and the jet trajectory can be approximated as a parabola. In order to accurately detect the maximum coverage area of the molten iron flow stream, the upper and lower boundary envelopes of the molten iron flow stream need to be captured. First, the Canny operator is used to extract the texture boundary of the molten iron flow video frame, and the image morphological opening operation is performed to remove the burr features of the contour image and the contour noise points inside the molten iron flow stream. The binary contour image is I binary (i,j). However, the molten iron flow contour boundary at this time is relatively rough and may have intermittent phenomena. Therefore, the Hough transform of the parabola equation is used to detect the molten iron flow contour binary image I binary (i,j) to obtain the upper and lower boundary maximum envelope equation (L1, L2) of the molten iron flow stream, and the closed coverage area I iron of the molten iron flow stream is intercepted through the upper and lower envelope equations, and the number of background pixels in the closed area is calculated:

[0083]

[0084] In the formula, [L1, L2] is the closed coverage area of the molten iron flow stream, I i,jrepresents the pixel value at (i, j) in the closed region, OTSU(I iron ) represents the threshold of the flow region foreground / background pixels, OTSU(I iron ) is the number of background pixels of the flow region, when the total number of background region pixels is greater than a set threshold, it indicates that there is an operator or toxic dust obstruction in the current frame, which cannot be used for molten iron flow state recognition and should be removed. Thus, the automatic recognition of the tapping and obstruction state can be realized, laying a foundation for intelligent perception of the molten iron flow state.

[0085] Step U3, image preprocessing and data set construction.

[0086] Step U3 mainly corrects the iron mouth position and angle, equalizes the scale and corrects the color transformation of the molten iron flow video frames collected in step U1 to ensure the consistency of the data set construction and the subsequent molten iron flow state monitoring, and prevents interference to the monitoring model caused by the installation position and distance difference of different tapping field equipment. Further, the molten iron flow video frames are pixel normalized, and the molten iron flow state data set is constructed by matching the time sequence labels with the labeled time stamps. The sub-function modules are described as follows:

[0087] Affected by the installation position, distance and angle of the video acquisition equipment, and the focal length, aperture size and exposure time of the imaging equipment, there are great differences in the starting position of the molten iron flow, the flow angle, the pixel scaling ratio and the pixel color space of the molten iron flow in the collected tapping process molten iron flow image frames, which brings great challenges to the universality of the molten iron flow state under different scenes. Therefore, the different tapping process molten iron flow images are corrected and standardized.

[0088] 1. Distortion and pose correction of molten iron flow image

[0089] First, the collected video frames are color converted and image enhanced, including image graying, image denoising and enhancement. The weighted average algorithm is used to realize the graying of the RGB image:

[0090] I gray = 0.3 * I R + 0.59 * I G + 0.11 * I B (3)

[0091] Due to the influence of factors such as installation position, distance and angle of the equipment for different tapping processes, the collected molten iron flow video frames are distorted, which has a certain influence on imaging and subsequent algorithm processing. In order to ensure the consistency of the imaging angle under different scenes, the distortion of the molten iron flow image needs to be corrected. As shown in Figure 3 ,Figure 3 In the formula, η is the normal vector of the plane where the molten iron flow is located, γ is the normal vector corresponding to the optical axis of the sensing device, O is the coordinate origin of the device installation position, and O' is the coordinate origin of the molten iron position. The camera internal parameters are obtained through camera calibration in a laboratory environment, and the camera external parameter matrix for the molten iron flow target is solved in combination with the on-site camera installation pose. The corresponding external parameter coordinate transformation matrix is as follows:

[0092] The rotation matrix around the X-axis direction relative to the camera relative to the taphole pose viewing angle a is as follows:

[0093]

[0094] The rotation matrix around the Y-axis is as follows:

[0095]

[0096] Since the device installation requirement in the Z-axis direction has no left and right tilt, there is no rotation, but there is a translation vector, so the transformation matrix is as follows:

[0097]

[0098] After the image is transformed, the equivalent viewing angle of the camera is perpendicular to the plane where the molten iron flow is located, and the lens optical axis passes through the initial position of the taphole, thereby ensuring the consistency of the imaging angle.

[0099] 2. Image equal-scale scaling of the molten iron flow

[0100] After distortion correction, the image is observed, and the uniform diameter of each taphole is d t The number of pixel points corresponding to the steel taphole casting in the image Y-axis direction is n t The actual physical size corresponding to a single pixel point in the real world is calculated as follows:

[0101]

[0102] To avoid the influence of the difference between the equivalent distances of the camera optical core and the taphole in different scenes on the subsequent algorithm, a standard scaling coefficient l s of the pixel point and the actual physical world is set, and the bicubic scaling algorithm is used to perform equal-scale scaling on the images in different scenes:

[0103]

[0104] In the formula, I o is the molten iron flow image after angle correction in different scenes, I s is the equal-scale scaled molten iron flow image, l / l s is the image equal-scale standard scaling coefficient, and B(.) is the bicubic image scaling algorithm.

[0105] After equal-scale scaling, the relative position and angle of the camera and the taphole remain constant under different scenarios, which greatly reduces the interference of the molten iron flow state recognition. However, the jet angle of the taphole in different tapping environments is different, so the initial angle of the flow needs to be further detected and transformed. Due to the difference in tapping direction, there is a difference between the left-to-right and right-to-left molten iron flow, so first, the right-to-left molten iron flow is horizontally mirrored:

[0106] I i,j width-i,j i < floor(width / 2) (9)

[0107] In the formula, I i,j is the pixel value corresponding to the coordinate (i,j), width is the image width, floor(·) is the floor function, and after mirror transformation, all the molten iron flow is left-to-right tapping.

[0108] 3. Flow ejection angle detection and transformation

[0109] After angle transformation and equal-scale scaling, due to the different design angles of each taphole, the angle of each taphole is different, and with the erosion of the taphole during tapping, the initial jet angle of the molten iron flow changes dynamically. In order to ensure the consistency of the initial conditions, the initial jet angle needs to be detected and standardized.

[0110] First, the OTSU algorithm is used to binarize the molten iron flow image frame to obtain the coverage area of the molten iron flow, and the coordinates of the taphole are searched:

[0111]

[0112] In the formula, I is the molten iron flow video frame, d t / l s is the number of Y-axis direction pixel points of each taphole after equal-scale scaling. The function S(.) realizes the search of the maximum continuous foreground pixel sequence of each column of the image frame, and is defined as:

[0113]

[0114] In the formula, T otsu is the binarization threshold of the OTSU algorithm, is the continuous foreground pixel sequence in the Y-axis direction. Thus, the initial horizontal coordinate i t ​Due to the constraint of the tundish, the flow in the initial jet direction vector is the same, the front section of the flow basically maintains the original speed direction, the middle and rear sections are affected by gravity and air resistance, and the flow rate presents a downward parabolic trajectory with a divergent state. Therefore, the front flow of the tundish is first detected, and the initial jet angle of the flow is further determined. First, define the Y-axis direction flow upper and lower boundaries:

[0115]

[0116] wherein, is the maximum continuous foreground pixel sequence of the i-th column of the flow image, is the initial element of the maximum continuous pixel sequence (i.e. the upper boundary of the flow), is the terminal element of the maximum continuous pixel sequence (i.e. the lower boundary of the flow). Through maximum value search and initial dimension constraint of the tundish, the abnormal recognition caused by background noise and flow shielding in the search process is ensured.

[0117] From the tundish position i t , the upper and lower boundaries of each column of the flow are stored in the upper and lower boundary linked list L t , b , and the search range is constrained by the divergence degree of the flow shape:

[0118]

[0119] wherein, Add(L, s) is a linked list insertion function, s is inserted into the linked list L, and ε is a flow diffusion coefficient, which is set to 1.1 to ensure accurate search of the initial flow section. The upper and lower boundaries of the initial flow section are linearly fitted to obtain the slope k t of the upper and lower boundaries of the initial flow section. b Then the final slope k of the initial flow section of the flow is:

[0120]

[0121] To ensure that the initial jet angle of the flow of different tundishes and different periods of the tundish process is the same, all flow image frames I of the tundish are corrected to a horizontal image I : o

[0122] I o = D·C·B·I (15)

[0123] wherein, D, C, and B are image translation matrix, image rotation matrix, and image translation matrix, respectively:

[0124] ​

[0125] The corrected image is taken as a reference for the initial position of the iron mouth, and the iron stream is uniformly scaled in the ROI region. The rectangular region coordinates of the cropped image are:

[0126]

[0127] In the formula, Rec((p x ,p y ),h,w) is a rectangular region description function, (p x ,p y ) is the starting coordinate of the upper left corner of the rectangular region, h and w are the height and width of the rectangular region respectively, and λ is the vertical and horizontal proportion coefficient of the iron stream:

[0128]

[0129] So far, the embodiment realizes the greying of the iron stream video frame, the field of view pose correction according to the camera installation pose in different scenes, and the stream scale. In order to accurately obtain the initial jet angle of the stream, the stream boundary extraction and jet angle calculation are realized, and finally the initial jet angle transformation and stream ROI extraction are realized.

[0130] 4. Image enhancement and data set construction

[0131] The recognition of the state of the iron stream mainly depends on the morphological characteristics of the stream edge. In order to further save the edge information of the iron stream and at the same time suppress the interference of the internal texture noise of the iron stream on the acquisition of the stream edge, bilateral filtering is used to realize the denoising and boundary preservation of the iron stream image, and the calculation formula is:

[0132]

[0133]

[0134] In the formula, q and p represent the input pixel point and the center pixel point respectively, i and j are the coordinates corresponding to the pixel point, and σs and σr are the control parameters of the pixel point space domain and value domain filtering respectively.

[0135] After image enhancement, each iron stream video frame is normalized to reduce the computational load of the subsequent neural network and further improve the model generalization ability. The standardization formula is:

[0136]

[0137] In the formula, μ is the mean of the image pixels, σ is the standard deviation of the image pixels, and I s is the normalized image.

[0138] The annotation form of the molten iron flow stream state in the tapping process is a list List(t, s), where (t, s) is the time and label group, and the stream state type includes normal, shielding, coke sticking, nozzle, and large flow. According to the time stamp of the molten iron flow video frame and the label group in the list, a complete data set is constructed.

[0139] Step U4, abnormal stream state skewness identification.

[0140] In the tapping process, molten iron is discharged from the furnace hearth through the tapping hole, and is easily affected by the damage of the mud bag, the change of the furnace hearth liquid level, and the blockage and erosion of the tapping hole, etc., to appear abnormal conditions such as nozzle and coke sticking, and the abnormal conditions can be directly reflected in the molten iron flow stream shape. Through the molten iron flow data set constructed in step U2, a ResNeXt network based on deep convolution is constructed for molten iron flow stream state recognition. The ResNeXt network forms a new residual module through input grouping convolution and fusion, realizing deep extraction of image features. However, in actual production process, the frequency of normal conditions is usually much higher than that of abnormal conditions, so the molten iron flow stream state data set presents a serious non-uniform state. In this case, the prediction of the molten iron flow stream state recognition model will be seriously biased towards the high-frequency label, resulting in missed reporting of abnormal conditions, seriously affecting the accuracy and reliability of the model.

[0141] To prevent the missed reporting of the molten iron flow stream abnormal state, first, the normal / abnormal state of the molten iron flow stream is realized to realize the detection of abnormal state. However, due to the serious non-uniform characteristics of the data of the normal / abnormal state of the molten iron flow stream, the detection of the abnormal state of the molten iron flow is difficult and the detection accuracy is low, so the focal loss function sensitive to sample classification accuracy is selected:

[0142] L fl =-(1-p t ) γ log(p t ) (22)

[0143] In the formula, γ is the modulation coefficient, p t is the probability of the stream state prediction as abnormal:

[0144]

[0145] In the formula, y is the label value of the current molten iron flow video frame, p is the prediction value of the neural network, and the label is 1 indicating the abnormal state of the molten iron flow stream and 0 indicating the normal state. However, this function cannot accurately increase the loss value of the missed reporting of abnormal conditions, therefore, the skewness loss function of the molten iron flow abnormality detection is constructed to increase the weight of the loss of abnormal samples in the network training process, and then the detection rate of abnormal samples is improved.

[0146]

[0147] In the formula, L represents the loss function of the samples during training, and p A and p N Let n represent the probabilities that the molten iron flow stream anomaly identification model predicts as either an abnormal or normal flow state. A n represents the number of samples in abnormal molten iron flow conditions. N This represents the number of samples in a normal state of the molten iron flow. Here, represents the first and second modulation coefficients of the loss function, and y represents the label value of the current molten iron flow video frame. y = 1 indicates an abnormal state of the molten iron flow, and y = 0 indicates a normal state. This function imposes a skewness constraint on the abnormal state of the molten iron flow by using the number of normal / abnormal samples and their probability differences. This shifts the model's learning focus away from abnormal samples, significantly increasing the false negative probability of abnormal samples while minimizing false positives of normal states.

[0148] During model training, the Adam optimizer was used to improve training speed, and the Dropout strategy was introduced to prevent overfitting and improve generalization performance. In actual training, the dataset was divided into training and test sets in an 8:2 ratio. The training ended when the detection rate of abnormal states in the molten iron flow stream in the test set met a set threshold C. h .

[0149] Step U5: Construction of the quasi-abnormal dataset.

[0150] Due to the effect of the skewed loss function in step U4, the loss value of outlier samples is strongly constrained during model training. Some normal samples that are similar to outliers are incorrectly identified as outliers. Therefore, quasi-outlier samples are defined as follows:

[0151]

[0152] In the formula, S is the quasi-anomaly judgment flag, where 1 indicates a quasi-anomaly and 0 indicates a non-quasi-anomaly sample. That is, the quasi-anomaly samples include real anomaly samples and normal samples that were incorrectly predicted as anomalies in step U4. For the constructed quasi-anomaly samples, the number of samples in each category are {n1, n2, ..., n}. m}, where m is the total number of quasi-abnormal sample categories, n1 is the number of normal samples incorrectly predicted as anomalous samples in step U4, and the remainder are the number of anomalous samples of each category. The dataset must satisfy the following:

[0153]

[0154] For quasi-exceptional samples, the high-precision detection of the abnormal state of the molten iron flow stream state is realized, but for accurate abnormal state recognition, the second stage of abnormal category recognition is still needed. For abnormal subcategory abnormal samples, the flow stream morphological characteristics have less distinction from normal samples, such as (coking, running large flow, etc.), and the category to which the abnormal state belongs needs to be judged through the morphological changes of the molten iron flow stream in a period of time. Therefore, for quasi-exceptional samples, a time sequence dynamic data set needs to be constructed, and the time sequence corresponding to a single sample is set as q image frames containing time in reverse order in the current video frame, wherein the sampling interval of each molten iron flow stream video frame remains consistent.

[0155] Step U6, flow stream dynamic feature extraction.

[0156] For the blast furnace tapping process, the size, shape and other characteristics of the molten iron flow stream are affected by multiple factors such as pressure fluctuation of the furnace, flowability of slag and iron, slag and iron level and real-time discharge rate, iron notch diameter and shape, etc. The characteristics of high similarity between categories and complex changes within categories, only relying on the original input of the image, have great interference and uncertainty. Therefore, the artificial experience is quantified in this embodiment, and the dynamic and steady-state characteristics of the molten iron flow stream are extracted for subsequent molten iron flow stream state recognition.

[0157] 1. Flow stream trajectory shape detection

[0158] The curve shape of the molten iron flow stream is an important reference index for judging the state of the molten iron flow. First, the Hough transform of the parabola fitting is used to detect the optimal parabola of the upper and lower boundary pixel sets L t ,L b of the molten iron flow stream in U3:

[0159]

[0160] In the formula, a t ,b t ,c t and a b ,b b ,c b are the coefficients of the upper and lower boundary parabola equations of the flow stream, the midpoint set M of the two parabolas is solved, and the parabola equation of the center axis of the molten iron flow stream is obtained by fitting the midpoint set M:

[0161] y m =a m x 2 +b m x+c m (28)

[0162] The three parabola equations perfectly describe the important morphological information such as the overall bending shape and divergence of the molten iron flow stream at different times.

[0163] 2. Stream boundary fluctuation rate and convex feature detection

[0164] During tapping, the normal stream boundary is usually smooth, and its shape and trend conform to the parabolic equation trend. However, when the tapping hole changes in shape due to erosion or blockage, or air is mixed into the stream of molten iron, the stream boundary of the stream of molten iron will exhibit abnormal shape characteristics such as convexity, fragmentation, and divergence. Defining and extracting the shape characteristics of the stream boundary is crucial for accurately identifying the stream state of the stream of molten iron. For the upper and lower boundary pixel sets L t ,L b and the upper and lower boundary parabolic equations y t ,y b , the stream boundary fluctuation rate V is defined as follows:

[0165]

[0166] wherein V represents the stream boundary fluctuation rate, λ represents the stream aspect ratio coefficient, d t represents the actual size of the steel nozzle at the tapping hole, l s represents the actual size of the object corresponding to a pixel point in the current video frame, L t (i) represents the corresponding y-coordinate of the upper boundary of the stream of molten iron at the horizontal coordinate i of the stream of molten iron image corresponding to the stream of molten iron video frame, y t (i) represents the corresponding value of the upper boundary parabolic equation of the stream of molten iron at the horizontal coordinate i of the stream of molten iron image, L b (i) represents the corresponding y-coordinate of the lower boundary of the stream of molten iron at the horizontal coordinate i of the stream of molten iron image, y b (i) represents the corresponding value of the lower boundary parabolic equation of the stream of molten iron at the horizontal coordinate i of the stream of molten iron image.

[0167] The stream boundary convexity feature also represents the divergence degree of the stream of molten iron. A stream convex boundary pixel coordinate set B = {b1, b2,..., b r} is constructed, wherein:

[0168]

[0169] wherein j represents the serial number of the stream convex boundary pixel horizontal coordinate set, r represents the number of boundary pixels contained in the convex boundary pixel horizontal coordinate set, L t (b j ) represents the corresponding y-coordinate of the upper boundary of the stream of molten iron at the horizontal coordinate b j of the stream of molten iron image corresponding to the stream of molten iron video frame, L b (b j ) represents the corresponding y-coordinate of the lower boundary of the stream of molten iron at the horizontal coordinate b j of the stream of molten iron image, y t (bj ) represents the x-axis of the molten iron flow graph. j At that point, the corresponding value of the parabolic equation of the upper boundary of the molten iron flow, y b (b j ) represents the x-axis of the molten iron flow graph. j The corresponding value of the equation of the parabola at the lower boundary of the molten iron flow.

[0170] Record the number N of the set of pixels with raised edges at the molten iron flow boundary in each image frame. B Simultaneously define the aspect ratio feature of the boundary bulge:

[0171]

[0172] Where R is the aspect ratio of the boundary convex feature, and l is the ratio of b1 to b2. r The equation of the straight line formed by the two boundary pixels is D. l (b j ,l) is b j The distance D from the corresponding boundary pixel to line l p (b1,b r ) for b1 and b r The Euclidean distance between the two boundary pixels corresponds to the boundary volatility and convexity characteristics of the flow stream. B ,R.

[0173] Step U7: Temporal state network construction and feature fusion.

[0174] For the quasi-anomaly dataset of the molten iron flow stream constructed in steps U5 and U6, data augmentation was performed in the time dimension to obtain complete time-series data. Combined with human experience, dynamic feature extraction of the molten iron flow stream was achieved. This embodiment first constructs a Conv_Lstm deep network model, whose input is the time series of the quasi-anomaly dataset of the molten iron flow stream, to extract the temporal features and image detail features of the molten iron flow stream. Simultaneously, an FNN network is constructed for training on the human-input features. At the end of the neural network, the outputs of the two network models are concatenated to achieve deep fusion of deep image features and human-input features. Finally, the network outputs the inference result of the current state of the molten iron flow stream, thus realizing online intelligent perception of the molten iron flow stream state.

[0175] Step U8: Data display and early warning.

[0176] In order to clearly and intuitively display the molten iron flow state recognition result of the system, facilitate the on-site operator to further evaluate and confirm the current recognition result, the real-time video of the high-speed sensor at the tapping hole is synchronously displayed on the software interface through the gigabit Ethernet, and the current state recognized by the system is displayed in the form of text. When the system recognizes that the current molten iron flow stream appears an abnormal state, the interface provides early warning information in a prominent color and text to attract the attention of the on-site operator. At the same time, the recognized molten iron flow state data is saved in the database, which is convenient for subsequent traceability analysis.

[0177] The method for intelligently perceiving the molten iron flow stream state in the blast furnace tapping process of the embodiment of the application uses a high-speed CMOS sensor to acquire high-speed video frames of the molten iron flow in the tapping process. By constructing a data set of different stream state categories and combining with a skew loss function, high-precision detection of the abnormal state of the molten iron flow stream is realized. At the same time, a dynamic construction of a quasi-abnormal data set and a stream dynamic / steady state feature extraction method are proposed, and the time sequence image sequence and the stream features are fused, and finally the online intelligent perception of the molten iron flow stream state is realized. The method and system have the advantages of high precision, strong stability, good real-time performance, low investment, convenient installation, long service life and the like, and fill the blank of the current monitoring of the molten iron flow stream state in the blast furnace tapping site.

[0178] Embodiment three

[0179] The specific implementation scheme of the application is further described in combination with the drawings. The embodiment of the application has been successfully applied in a 1# blast furnace of a domestic steel plant. In order to ensure the efficient and stable operation of the algorithm, one high-performance computer is required, and the gigabit network is connected with the imaging equipment of the blast furnace tapping hole. Next, the specific implementation steps are as follows: one high-performance computer is installed in the blast furnace control room, and a stable gigabit Ethernet network is configured.

[0180] Step S301, high-speed acquisition of the molten iron flow. The molten iron flow image acquisition equipment is directly connected through the gigabit Ethernet to realize high-speed real-time acquisition of the molten iron flow image.

[0181] Step S302, recognition of the molten iron flow area shielding state. The open / tight tapping hole state and the stream shielding area are automatically recognized to determine whether the molten iron flow stream state recognition algorithm is implemented at present, realize the self-adaptation of the molten iron flow state recognition algorithm, greatly save the computing resources, and at the same time improve the credibility of the molten iron flow state recognition result.

[0182] Step S303, image preprocessing and data set construction. Due to the large difference in the installation distance, pose, perspective and scale of the molten iron flow of different equipment, through the equal-scale correction of the distance and pose, the efficient transformation of the image scene in different scenes is realized, and a high-quality data set is constructed.

[0183] Step S304, abnormal stream state skewness identification. First, a Res_Next network is constructed, and through the strong constraint of the skewness loss function on the abnormal samples, efficient detection of abnormal samples is realized, greatly preventing the false negatives of abnormal stream state.

[0184] Step S305, quasi-abnormal data set construction. For the prediction results of step S304, to further identify the abnormal category to which the abnormal state belongs, a quasi-abnormal data set of molten iron stream is further constructed.

[0185] Step S306, stream dynamic and steady-state feature extraction. According to artificial experience, important molten iron stream features are defined, laying a foundation for subsequent accurate identification of molten iron stream abnormal categories.

[0186] Step S307, time series state and stream feature fusion. The Conv_Lstm neural network and the FNN network are constructed in this embodiment to be used for deep feature extraction of time series molten iron stream image sequences and training of stream features, respectively, and then through further fusion of deep features of the two networks, accurate identification of the final molten iron stream state is realized.

[0187] Step S308, data display and storage. The real-time picture of the molten iron stream in the tapping process is displayed in the software interface, and the molten iron stream state recognition result is displayed in real time and the abnormal state warning is provided.

[0188] Reference Figure 4 The online intelligent perception system for the molten iron stream state of the blast furnace provided in the embodiment of the present application comprises a memory 10, a processor 20, and a computer program stored in the memory 10 and capable of running on the processor 20, wherein the processor 20 implements the steps of the online intelligent perception method for the molten iron stream state of the blast furnace provided in the embodiment of the present application when executing the computer program.

[0189] The specific working process and working principle of the online intelligent perception system for the molten iron stream state of the blast furnace in the embodiment of the present application can refer to the working process and working principle of the online intelligent perception method for the molten iron stream state of the blast furnace in the embodiment of the present application.

[0190] The above is only the preferred embodiment of the present application and is not used to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for online intelligent perception of molten iron stream flow state of a blast furnace, characterized in that, The method comprises: Collecting the video frames of the molten iron flow in the tapping process and constructing the molten iron flow stream state dataset of different stream state categories, wherein the collecting the video frames of the molten iron flow in the tapping process comprises: Identifying the tapping state of the tapping hole, and the specific identification formula is: Wherein, T is the tapping state of the tapping hole, 1 represents that the tapping is in progress, 0 represents that the tapping is not in progress, I represents the video frame of the molten iron flow, max(I) represents the maximum value of the pixels in the video frame of the molten iron flow, OTSU(I) represents the image threshold value of the OTSU algorithm, sum(I>OTSU(I)) represents the number of pixels in the image that exceeds the threshold value, ω, ξ, τ respectively represent the first threshold value parameter, the second threshold value parameter and the third threshold value parameter of the self-defined tapping state detection; According to the identification result of the tapping state of the tapping hole, the video frames of the molten iron flow in the tapping process are collected; Constructing a ResNeXt network based on deep convolution to obtain a molten iron flow stream abnormal state recognition model, and detecting the molten iron flow stream abnormal state according to the molten iron flow stream abnormal state recognition model, wherein the skewness loss function of the molten iron flow stream abnormal state recognition model is specifically: wherein L represents the loss function of the sample in the training process, p A and p N are the probabilities of the molten iron stream abnormal state identification model being predicted as an abnormal stream state and a normal stream state, respectively, n A is the number of molten iron stream abnormal state samples, n N is the number of molten iron stream normal state samples, l and θ are the first and second modulation coefficients of the loss function, respectively, and y is the label value of the current molten iron stream video frame, and y = 1 represents the molten iron stream abnormal state and y = 0 represents the molten iron stream normal state. According to the identification result of the molten iron flow stream abnormal state, constructing a quasi-abnormal time series dataset; Extracting the stream dynamic features of the video frames of the molten iron flow; Using the Conv_Lstm network to extract the high-dimensional features of the quasi-abnormal time series dataset, using the FNN network to extract the high-dimensional features of the stream dynamic features, and splicing and fusing the high-dimensional features extracted by the two networks to obtain a molten iron flow stream state online intelligent perception model, and detecting the molten iron flow stream state according to the molten iron flow stream state online intelligent perception model.

2. The method for online intelligent perception of the flow state of the molten iron stream of a blast furnace according to claim 1, characterized in that, After collecting the video frames of the molten iron flow in the tapping process, it further comprises: Performing angle transformation and equal-scale scaling on the video frames of the molten iron flow.

3. The method for online intelligent perception of the flow state of the molten iron stream of a blast furnace according to claim 2, characterized in that, According to the identification result of the molten iron flow stream abnormal state, constructing a quasi-abnormal time series dataset comprises: According to the identification result of the molten iron flow stream abnormal state, obtaining quasi-abnormal samples, wherein the quasi-abnormal samples include real abnormal samples identified as abnormal samples and normal samples incorrectly identified as abnormal samples; According to the time series corresponding to the samples in the quasi-abnormal samples, a quasi-abnormal time series dataset is obtained, wherein the time series corresponding to a single sample is specifically a preset number of image frames contained in the current video frame in reverse order of time.

4. The method for online intelligent perception of the flow state of the molten iron stream of a blast furnace according to claim 3, characterized in that, Extracting the stream dynamic features of the video frames of the molten iron flow comprises: Extracting the stream trajectory of the video frames of the molten iron flow; Obtaining the stream boundary fluctuation rate and / or convex feature of the stream trajectory as the stream dynamic features of the video frames of the molten iron flow.

5. The method for online intelligent perception of the flow state of the molten iron stream of a blast furnace according to claim 4, characterized in that, The specific formula for obtaining the stream boundary fluctuation rate of the stream trajectory is: where V represents the flow boundary fluctuation rate, λ represents the flow aspect ratio coefficient, d t represents the actual size of the steel nozzle at the taphole, l s represents the actual size of a pixel point of the current video frame corresponding to the photographed object, L t (i) represents the corresponding ordinate of the upper boundary of the molten metal flow at the horizontal coordinate i of the molten metal flow image corresponding to the molten metal flow video frame, y t (i) represents the corresponding value of the parabolic equation of the upper boundary of the molten metal flow at the horizontal coordinate i of the molten metal flow image, L b (i) represents the corresponding ordinate of the lower boundary of the molten metal flow at the horizontal coordinate i of the molten metal flow image, y b (i) represents the corresponding value of the parabolic equation of the lower boundary of the molten metal flow at the horizontal coordinate i of the molten metal flow image.

6. The method for online intelligent perception of the flow state of the molten iron stream of a blast furnace according to claim 5, characterized in that, Obtaining the convex feature of the stream trajectory comprises: constructing a set of stream protrusion boundary pixel abscissas B = {b1, b2,..., b r}, wherein: wherein j represents the serial number of the set of lateral coordinates of the convex boundary pixels of the streamer, r represents the number of boundary pixels contained in the set of lateral coordinates of the convex boundary pixels, L t (b j ) represents the corresponding longitudinal coordinate of the upper boundary of the streamer at the lateral coordinate b j of the stream image corresponding to the streamer video frame, L b (b j ) represents the corresponding longitudinal coordinate of the lower boundary of the streamer at the lateral coordinate b j of the stream image, y t (b j ) represents the corresponding value of the parabolic equation of the upper boundary of the streamer at the lateral coordinate b j of the stream image, y b (b j ) represents the corresponding value of the parabolic equation of the lower boundary of the streamer at the lateral coordinate b j of the stream image. Recording the number of sets of molten iron flow boundary convex pixels in the video frames of the molten iron flow, and calculating the boundary convex aspect ratio feature, and the specific calculation formula is: Where R is the aspect ratio of the boundary convex feature, and l is the ratio of b1 to b2. r The equation of the straight line formed by the two boundary pixels is D. l (b j ,l) is b j The distance D from the corresponding boundary pixel to line l p (b1,b r ) for b1 and b r The Euclidean distance between the two boundary pixels.

7. A molten iron flow stream state online intelligent perception system for a blast furnace, the system comprising: A memory (10), a processor (20), and a computer program stored on the memory (10) and executable on the processor (20), wherein the processor (20) implements the steps of the method of any one of claims 1 to 6 when executing the computer program.

Images