Method for determining optimal cross-section image of sintering machine tail and method for predicting FeO content of sintered ore based on same

Abstract

According to the method for determining the optimal cross-section image of the sintering machine tail and the method for predicting the FeO content of the sintering ore based on the optimal cross-section image of the sintering machine tail, firstly, based on the speed of a sintering pallet, the position of a sintering end point and the moment when the sintering ore falls into a chute of historical data, the optimal cross-section image of the sintering machine tail is obtained; carrying out linear regression to obtain sintered ore fracture starting moments corresponding to different sintering pallet speeds and different sintering end point positions; then, the actual breakage starting moment of the sinter is determined; and thirdly, determining a camera triggering acquisition moment and an acquisition ending moment, and finally, determining a target time period, and determining frames in the target time period as optimal cross-section image frames. According to the method for predicting the FeO content of the sintered ore based on the method for determining the optimal section of the sintering machine tail, a convolutional neural network is used as a basic prediction model, meanwhile, the structure is set as a residual network structure, and an optimal section image is used as the input of the network.

Description

Technical Field

[0001] The present invention belongs to the technical field of iron ore sintering production, and particularly relates to a method for determining the optimal cross-sectional image at the tail of a sintering machine and a method for predicting the FeO content of sinter based on this. Background Technique

[0002] The quality indicators of sinter mainly include chemical composition, physical properties and metallurgical properties. Among them, the FeO content is an important indicator in the chemical composition of sinter. It is closely related to the drum strength and reduction performance of sinter, and at the same time reflects the heat control level in the sintering process. Therefore, the FeO content is a comprehensive indicator for measuring the quality of sinter and the production operation level. In addition, the FeO content of sinter is an important factor affecting blast furnace smelting. Some studies have shown that when the FeO content of the burden changes by 1%, the coke ratio and molten iron output of the blast furnace will change by 1% - 1.5%. At present, the detection methods for the FeO content of sinter mainly fall into direct measurement and indirect measurement. Direct measurement includes chemical composition analysis, magnetic permeability analysis, etc., and indirect measurement is mainly model prediction based on data driving. Chemical composition analysis is the simplest, most direct and accurate measurement method, and it is also the most commonly used method in the current production site. However, since sintering is a continuous production process, sampling and analyzing sinter requires manual participation and has a long cycle (4 - 6 hours), which is difficult to meet the requirements of production process control. For model prediction based on data driving, it will present various technical considerations according to different conceived technical structures. And regardless of the model based on which theoretical structure, it requires an image as input, and how to capture the optimal cross-sectional image at the tail of the machine is an objective problem that the industry has been looking for corresponding solutions.

[0003] The invention application with the application number: CN201010597200.1 discloses "an automatic capture method for an effective image of the red cross-section at the tail of a sintering machine", including (1) image acquisition; (2) image inspection trigger; (3) image preprocessing; (4) image validity inspection; among which, the image acquisition method in step (1) is to continuously acquire real-time images of the red cross-section at the tail of the sintering machine through an image acquisition device; the image inspection trigger in step (2) is to extract the real-time images of the red cross-section at the tail of the machine acquired at a certain time period, and then create an independent analysis thread for each target image; the image preprocessing in step (3) is to perform grayscale conversion and red light RGB image conversion on the acquired original image. The red light RGB image conversion means taking the G value and B value of the RGB image of the original image as zero and only retaining the RGB image of the R value; the image validity inspection in step (4) includes image validity feature extraction and image validity feature analysis.

[0004] The invention application with the application number: CN202010469161.0 discloses "a method and system for obtaining the image of the tail section of a sintering machine", including: installing a light source on the tail platform of the sintering machine to illuminate the fracture area of the sintered cake at the tail of the sintering machine; making the camera follow the flipping of the tail trolley of the sintering machine so that the shooting axis of the camera is on the same straight line as the extension axis of the sintered cake on the tail trolley; collecting the time when the fractured sintered cake drops and vibrates; collecting the image at a predetermined moment before the time when the sintered cake drops and vibrates from the image taken by the camera as an alternative image of the tail section of the sintered cake.

[0005] The invention application with the application number: CN202111075777.0 discloses "a method and system for obtaining the best cross-section image of the sintering machine tail", by determining the preset area where the cross-section of the sintering machine tail is located, using an image acquisition device and a temperature measurement device to continuously collect real-time image information of the preset area at the same time, the image information includes the average temperature and the acquisition time of the preset area, using the historical trend curve of the average temperature of the preset area with respect to the acquisition time to determine the best acquisition time point, and then extracting the corresponding best cross-section image according to the best acquisition time point.

[0006] The invention application with the application number: CN202210374244.0 discloses "a method, system, device and medium for processing the image of the tail section of a sintering machine", and its steps are: obtaining the target image of the tail section of the sintering machine to be processed; inputting the target image of the tail section of the sintering machine into the pre-trained Attention-Dense-Resnet U-net neural network; receiving the target area calibration result of the target image of the tail section of the sintering machine output by the Attention-Dense-Resnet U-net neural network.

[0007] The invention application with the application number: CN202211499407.4 discloses "a method for capturing key frames of the tail section of a partition-triggered sintering machine", including the following steps: Step 1, collecting the video of the cross-section of the tail section of the sintering machine; Step 2, manually dividing the collected video window into two functional areas; Step 3, monitoring the average temperature Tr1 of the R1 area to judge whether the sintered ore of the previous trolley is about to drop; Step 4, if the sintered ore of the previous trolley is about to drop, sending an instruction to trigger the temperature monitoring of the R2 area; Step 5, monitoring the average temperature Tr2 of the R2 area to judge whether the dust raised by the dropping of the sintered ore is about to block the line of sight; Step 6, if the dust raised by the dropping of the sintered ore is about to block the line of sight, sending an instruction to trigger the capture of the key frame.

[0008] The paper titled "Research on Key Technologies of Real-time Prediction System for FeO Content in Sinter" by Chongqing University, based on the idea of difference, determines the best cross-section image by finding the moment t with the largest difference within a cycle, and the corresponding cross-section image is the best cross-section image.

[0009] The invention application with the application number CN201910642094.5 discloses "A Method and System for Detecting FeO Content in Sinter". This method obtains thermal images, extracts key frame images by combining the change law of dust at the tail of the sintering machine, and based on the key frame images, uses the geometric features of the tail trolley to extract the infrared thermal images of interest, thereby obtaining the infrared thermal images of the sinter cross-section. Based on the infrared thermal images of the sinter cross-section, shallow and deep features describing the quality of the sinter are extracted, a multi-phase thermodynamics model of the sintering process based on the Gibbs free energy theorem is established, and according to the multi-phase thermodynamics model, the classification features of the FeO content at the highest temperature of the sinter are obtained and a prediction model of the FeO content based on multi-class heterogeneous features is established. Then, using the shallow features, deep features, and FeO content classification features, the FeO content of the sinter is predicted in real-time online.

[0010] The invention application with the application number CN202110320721.0 discloses "A Method and System for Predicting FeO in Sinter by Data and Knowledge Fusion". By obtaining the sample parameters associated with the FeO content in the sinter and the data density center of the sample parameters, based on the highest temperature of the sintering material layer, the FeO content grade of the sinter is inferred online using the mechanism knowledge base, and according to the data density center of the sample parameters and the FeO content grade of the sinter, an online estimation model of the FeO content in the sintering process is constructed to achieve the prediction of the FeO content in the sinter, solving the technical problem of low prediction accuracy of the existing FeO content in the sinter. Moreover, by using the extraction of the data density center based on the kernel function high-dimensional mapping to solve the problem of data inconsistency caused by different sampling frequencies, it is beneficial to improve the prediction accuracy of the FeO content in the sinter.

[0011] The invention application with the application number CN202110497744.9 discloses "Construction and Application of Soft Measurement Model for FeO Content in RVM Sinter". It combines the sensor data in the sintering process with the cross-section image data at the tail of the sintering machine to jointly complete the construction of the model, making the data information more abundant. In addition, feature construction is carried out on the data to make it have a stronger explanatory ability for the non-linearity and dynamics of the sintering data. On this basis, an RVM model is constructed to enable it to have an accurate online soft measurement ability for the FeO content in the complex sintering process. Summary of the Invention

[0012] The present invention provides a method for determining the best cross-section image at the tail of the sintering machine and a prediction method for the FeO content in the sinter based on this. The technical solution is as follows:

[0013] A method for determining the optimal cross-section image at the tail of a sintering machine, comprising the following steps:

[0014] S1: Based on the sintering trolley speed, sintering end position, and the moment when sintered ore falls into the chute in historical data, statistically analyze and linearly regress the moment when the sintered ore starts to break corresponding to different sintering trolley speeds and different sintering end positions;

[0015] S2: According to step S1, combined with the current sintering trolley speed and the current sintering end position, determine the moment when the current sintered ore starts to break at the tail of the sintering machine;

[0016] S3: Trigger the camera to collect image frames at a set frequency at a set moment before the current sintered ore reaches the moment of starting to break, until the camera ends the collection when the sintered ore falls into the chute, and determine the target period according to the brightness change rule of the camera field of view during the period from the start of the sintered ore breaking until the sintered ore falls and generates dust;

[0017] S4: Determine the frames within the target period as the optimal cross-section image frames.

[0018] Furthermore,

[0019] Step S1 is specifically as follows:

[0020] SS1: According to the sintering trolley speed in historical data, complete the classification establishment of the sintered ore trolley speed;

[0021] SS2: For the sintering trolley speeds under each category, based on the sintering end position in historical data and the moment when the sintered ore falls into the chute, statistically analyze and linearly regress the corresponding relationship between the sintering end position and the moment when the sintered ore starts to break.

[0022] Furthermore,

[0023] Regarding "determine the target period according to the brightness change rule of the camera field of view during the period from the start of the sintered ore breaking until the sintered ore falls and generates dust" in step S3, specifically: Determine the third mutation moment of the camera field of view as the starting moment of the target period, and determine the period from this moment to the end of the camera collection moment as the target period.

[0024] Furthermore,

[0025] The optimal cross-section image frames in step S4 can be any frame within the target period; or determine the optimal cross-section image frames within the target period according to the overall plane temperature.

[0026] A method for predicting the FeO content of sintered ore based on the method for determining the optimal cross-section image at the tail of a sintering machine

[0027] Taking a convolutional neural network as a prediction model and using the best cross-sectional image as input to complete the prediction.

[0028] Furthermore,

[0029] The convolutional neural network is a residual network.

[0030] Furthermore,

[0031] Add Gaussian noise with a set variance to the output of each residual mapping.

[0032] A method for determining the best cross-sectional image at the tail of a sintering machine and a prediction method for the FeO content of sinter ore based on this. In the method for determining the best cross-sectional image at the tail of a sintering machine, historical data is fully utilized and combined with BTP (sintering end point) to realize the determination of the relationship conversion at the start time of sinter ore fracture, so that the start time of sinter fracture can be determined according to different sintering trolley speeds and different sintering end points. Then, based on the start time of sinter ore fracture and the brightness change law in the camera field of view during the period from the start time of sinter ore fracture to the dropping time, the determination of the best cross-sectional image frame at the machine tail is established. In the prediction method for the FeO content of sinter ore based on this, the basic purpose is prediction accuracy, and the robustness and generalization ability of the model are also considered. Among them, multiple attentions are paid from the construction of the network structure of the model to the step size setting of downsampling in the convolutional layer and pooling layer to ensure the generalization ability of the model; to ensure prediction accuracy and save data preprocessing, a convolutional neural network is used. Considering the affine transformation and non-linear transformation layer by layer, and the affine transformation is based on matrix multiplication operation, so the model further adopts the structure form of a residual network, changing multiplication to addition, thereby improving the stability and ease of training and suppressing the attenuation of the gradient. In terms of robustness, the response is comprehensively considered and established from the perspective of ensuring the input accuracy and model regularization. Ensuring the input accuracy refers to obtaining the best cross-sectional image in the technical solution and using it as input; from the perspective of regularization, Gaussian noise with a set variance is added to the output of each residual mapping, and it is proved by practice that the effect is better than L1 or L2 regularization. Description of the Drawings

[0033] Figure 1 It is a schematic diagram of the steps for determining the best cross-sectional image at the tail of a sintering machine in the present invention. Detailed Embodiments

[0034] Next, a method for determining the best cross-sectional image at the tail of a sintering machine and a prediction method for the FeO content of sinter ore based on this in the present invention will be further specifically described according to the drawings in the specification and the detailed embodiments.

[0035] Design Principle:

[0036] Determination part of the optimal cross-sectional image at the tail of the sintering machine:

[0037] First, the calculation module of the controller reads the corresponding sintering end point information and the information of the time point when the sinter ore falls and hits the chute at the same sintering trolley speed from the historical data; forms statistical information, and then based on the statistical information, combines the following formula to linearly regress the corresponding relationship between the sintering end point position and the moment when the sinter ore starts to break,

[0038]

[0039] In the above formula,

[0040] t0: The moment when the sinter ore falls into the chute, unit: s;

[0041] S: The length of the sintering trolley, unit: m;

[0042] V: The speed of the sintering trolley, unit: m / min;

[0043] α: Strength influence factor, related to the sintering end point;

[0044] t: Unit time quantity, unit: s;

[0045] t1: The moment when the sinter ore completely breaks and falls into the chute from the start of breaking, which is an empirical constant.

[0046] Among them, αt is the time period from when the sinter ore reaches the tail of the machine to the start of breaking, which is related to the sintering end point. The linear regression is based on the least squares method.

[0047] Secondly, based on the above relationships, combined with the actual trolley speed and the position of the sintering end point, the fracture start time of the current sintered ore can be calculated. The set forward shift period (the so-called forward shift period refers to the period set forward from the fracture start time of the sintered ore) is used as the trigger signal to trigger the camera to collect image frames at a set frequency. After receiving the trigger signal, the camera collects image frames until the sintering furnace falls to the chute to trigger the camera to end the collection. Then, the computer determines the target period according to the brightness change law of the camera's field of view during the period from the start of the fracture of the sintered ore until the sintered ore falls and raises dust. Here, the characteristics that the appearance of cracks in the sintered ore will cause the first brightness mutation in the camera's field of view, the appearance of the cross-section will cause the second brightness mutation in the camera's field of view, the complete formation of the cross-section will cause the third brightness mutation in the camera's field of view, and the raising of dust when the sintered ore falls will cause the fourth brightness mutation in the camera's field of view are utilized. The moment corresponding to the third brightness mutation of the camera's field of view is determined as the start time of the target period, and the period from this moment to the end of the camera's collection is determined as the target period. Thus, the determination of the target period is completed. Then, the best cross-section image can be any frame within the target period, which can already meet the target requirements; or, in order to pursue higher precision, the best cross-section image frame within the target period can be determined according to the overall plane temperature, but the significance and result of this pursuit are not so prominent.

[0048] The camera mentioned in this technical solution refers to a thermal imager.

[0049] Prediction part of the FeO content of sintered ore based on the best cross-section image at the tail of the sintering machine:

[0050] First, in order to reduce data preprocessing and ensure prediction accuracy, a convolutional neural network model that is more friendly to image prediction is adopted. Secondly, to ensure the robustness of the model, the technical solution adopts a technical structure setting of input accuracy guarantee + regularization idea. For the input accuracy guarantee, the best cross-section image determined by this technical solution is used as the input of the model. The embodiment of the regularization idea is reflected in adding Gaussian noise with a set variance to the output of each residual mapping. In order to ensure that the gradient does not decay excessively, and considering that the basic calculation steps of the convolutional neural network are to perform affine transformation and non-linear transformation layer by layer (corresponding to convolution pooling and activation respectively), and the affine transformation is based on matrix multiplication, the model further adopts the structure form of a residual network, changing multiplication to addition, thereby improving the stability and ease of training and suppressing the decay of the gradient. At the same time, the model also makes the following attentions and settings:

[0051] 1. The convolutional layer of the improved convolutional neural network consists of three 3×3 convolutional kernels. The stride of the first convolutional layer is set to (1, 2) to ensure that the network can learn more deep features in the height direction of the sintered ore layer section.

[0052] 2. In the downsampling block of the improved convolutional neural network, the stride of the 1×1 convolutional layer in the right channel is set to 1, and a 2×2 average pooling layer with a stride of 2 is added before it to avoid ignoring the cross-sectional image information.

[0053] 3. The SELU function is used as the activation function to ensure that the gradient during the training process will not explode or vanish.

[0054] The constructed deep residual network ResNet-18 consists of a max pooling layer, two residual blocks, downsampling block + residual block, downsampling block + residual block, downsampling block + residual block, and a global average pooling layer.

[0055] Next, it will be further elaborated in the form of embodiments to form a more comprehensive understanding of the above elaboration and at the same time verify the accuracy of the prediction model of this solution.

[0056] Embodiment

[0057] First, according to the speed of the current sintering trolley and the sintering end point, using the optimal cross-section determination method determined by this technical solution, the optimal cross-sectional image is obtained as the input of the model.

[0058] Then, the deep residual learning architecture (residual network, ResNet-18) is selected as the basic framework of the FeO measurement model. The expression of the residual function is F(x) = H(x) - x, and the output of the residual block is as follows:

[0059] y = F(x, {W i}) + x,

[0060] In the formula, x and y are the input and output vectors of the residual block, the function F(x, {W i}) is the residual mapping to be learned, and W i is the weight to be learned in the convolutional layer.

[0061] Again, three 3×3 convolutional kernels are used to replace the 7×7 convolutional kernel of the baseline network (ResNet-18) to reduce the risk of model overfitting; the stride of the first convolutional layer in the input layer of the baseline network is changed to (1, 2) to ensure that the network can learn more deep features in the height direction of the image; the stride of the 1×1 convolutional layer in the right channel of the downsampling block of the baseline network is changed to 1, and a 2×2 average pooling layer with a stride of 2 is added before it to avoid ignoring information and at the same time have a small impact on the model complexity.

[0062] Finally, the obtained dataset composed of the input is randomly divided into a training set and a test set in a ratio of 7:3, and the model is trained and a corresponding FeO measurement system is established. Adam is used as the optimizer of the model, the Huber loss with δ = 0.2 is used as the loss function of the model, SELU is selected as the activation function of the network, and the weight matrix of each layer of the model is initialized using the Xavier method. The training set is fed into the network for training, and the method of preheating the training model is used to accelerate the convergence of the model to achieve the training effect.

[0063] To prove the better detection performance of this model, the present invention selects the traditional VGG-16 and Resnet-18 models for comparative experiments, and finally conducts comparative experiments on the two models with the same training set and test set. Table 1 calculates and statistically analyzes the relevant parameters of the measurement errors of these three types of models, where the measurement accuracy is judged based on ±0.5%.

[0064] Table 1 Measurement Error Statistics

[0065] Model Root Mean Square Error Average Relative Error (%) Measurement Accuracy Rate (%) VGG-16 0.49 4.35 71.4 ResNet-18 0.36 3.13 82.4 The method of the present invention 0.29 2.50 91.2

Claims

1. A method for determining the best cross-sectional image of a sintering machine tail, characterized in that The steps include: S1: Based on the historical data of sintering trolley speed, sintering end position and the time when the sintered ore falls into the chute, statistical analysis and linear regression are performed to determine the sintered ore fracture start time corresponding to different sintering trolley speeds and different sintering end positions; S2: According to step S1, the current sintering trolley speed and the current sintering end position are combined to determine the starting fracture moment of the current sintered ore at the tail end of the sintering machine; S3: At a set time before the current sinter reaches the start of fracture, the camera is triggered to capture image frames at a set frequency. The camera is triggered to end the capture when the sinter falls into the chute. The target period is determined based on the brightness change pattern of the camera field of view from the start of fracture to the fall of the sinter to generate dust. S4: Determine the frames within the target time period as the optimal cross-sectional image frames.

2. The method for determining the optimal cross-sectional image of the sintering machine tail according to claim 1, characterized in that: Step S1 is specifically as follows: SS1: Complete the classification of sintering trolley speeds based on the sintering trolley speeds in the historical data; SS2: For each type of sintering trolley speed, based on the historical data of the sintering end point position and the time when the sintered ore falls into the chute, statistical analysis and linear regression are used to determine the corresponding relationship between the sintering end point position and the time when the sintered ore begins to break.

3. The method for determining the optimal cross-sectional image of the sintering machine tail according to claim 1, characterized in that: The "target period is determined based on the brightness change pattern of the camera's field of view from the time the sintered ore begins to fracture until the sintered ore falls and generates dust" described in step S3 is specifically as follows: the moment of the third sudden change in the camera's field of view is determined as the starting moment of the target period, and the period from this moment to the moment when the camera ends collecting data is determined as the target period.

4. The method for determining the optimal cross-sectional image of a sintering machine tail according to claim 1, characterized in that: The best cross-sectional image frame in step S4 may be any frame within the target time period; or the best cross-sectional image frame within the target time period may be determined based on the overall plane temperature.

5. A method for predicting the FeO content of sintered ore based on the method according to claim 1, characterized in that: The convolutional neural network is used as the prediction model and the optimal cross-sectional image is used as input to complete the prediction.

6. The method for predicting the FeO content of sintered ore according to claim 5, wherein: The convolutional neural network is a residual network.

7. The method for predicting the FeO content of sintered ore according to claim 6, wherein: Gaussian noise with a set variance is added to the output of each residual map.

Images