CN113297886A - Material surface ignition effect detection method and device based on convolutional neural network - Google Patents
Material surface ignition effect detection method and device based on convolutional neural network Download PDFInfo
- Publication number
- CN113297886A CN113297886A CN202010797640.5A CN202010797640A CN113297886A CN 113297886 A CN113297886 A CN 113297886A CN 202010797640 A CN202010797640 A CN 202010797640A CN 113297886 A CN113297886 A CN 113297886A
- Authority
- CN
- China
- Prior art keywords
- charge level
- neural network
- image
- convolutional neural
- images
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 36
- 230000000694 effects Effects 0.000 title claims abstract description 25
- 238000001514 detection method Methods 0.000 title claims abstract description 22
- 239000000463 material Substances 0.000 title description 14
- 238000012360 testing method Methods 0.000 claims abstract description 27
- 238000000034 method Methods 0.000 claims description 56
- 238000012549 training Methods 0.000 claims description 36
- 238000013528 artificial neural network Methods 0.000 claims description 26
- 230000008569 process Effects 0.000 claims description 26
- 208000012868 Overgrowth Diseases 0.000 claims description 18
- 238000007781 pre-processing Methods 0.000 claims description 13
- 238000004458 analytical method Methods 0.000 claims description 11
- 238000002347 injection Methods 0.000 claims description 7
- 239000007924 injection Substances 0.000 claims description 7
- 238000004364 calculation method Methods 0.000 claims description 6
- 238000011176 pooling Methods 0.000 claims description 4
- 238000011478 gradient descent method Methods 0.000 claims description 3
- 230000002159 abnormal effect Effects 0.000 abstract description 10
- 238000002844 melting Methods 0.000 abstract description 9
- 230000008018 melting Effects 0.000 abstract description 9
- 238000003062 neural network model Methods 0.000 abstract description 5
- 230000006870 function Effects 0.000 description 16
- 238000005245 sintering Methods 0.000 description 11
- 238000010586 diagram Methods 0.000 description 7
- XEEYBQQBJWHFJM-UHFFFAOYSA-N Iron Chemical compound [Fe] XEEYBQQBJWHFJM-UHFFFAOYSA-N 0.000 description 6
- 239000007789 gas Substances 0.000 description 5
- 239000011159 matrix material Substances 0.000 description 5
- 238000012545 processing Methods 0.000 description 5
- 238000004422 calculation algorithm Methods 0.000 description 4
- 239000000203 mixture Substances 0.000 description 4
- 230000009467 reduction Effects 0.000 description 4
- 238000004590 computer program Methods 0.000 description 3
- 239000000446 fuel Substances 0.000 description 3
- 229910052742 iron Inorganic materials 0.000 description 3
- OKTJSMMVPCPJKN-UHFFFAOYSA-N Carbon Chemical compound [C] OKTJSMMVPCPJKN-UHFFFAOYSA-N 0.000 description 2
- 229910052799 carbon Inorganic materials 0.000 description 2
- 238000004891 communication Methods 0.000 description 2
- 230000003247 decreasing effect Effects 0.000 description 2
- 230000002950 deficient Effects 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 238000001914 filtration Methods 0.000 description 2
- 238000010304 firing Methods 0.000 description 2
- 238000012417 linear regression Methods 0.000 description 2
- 238000011068 loading method Methods 0.000 description 2
- 238000012544 monitoring process Methods 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 239000002245 particle Substances 0.000 description 2
- 239000002994 raw material Substances 0.000 description 2
- 239000007787 solid Substances 0.000 description 2
- 238000003860 storage Methods 0.000 description 2
- 230000000007 visual effect Effects 0.000 description 2
- 230000006978 adaptation Effects 0.000 description 1
- QVGXLLKOCUKJST-UHFFFAOYSA-N atomic oxygen Chemical compound [O] QVGXLLKOCUKJST-UHFFFAOYSA-N 0.000 description 1
- 238000013145 classification model Methods 0.000 description 1
- 230000006835 compression Effects 0.000 description 1
- 238000007906 compression Methods 0.000 description 1
- 239000000428 dust Substances 0.000 description 1
- 230000004907 flux Effects 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000002156 mixing Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 229910052760 oxygen Inorganic materials 0.000 description 1
- 239000001301 oxygen Substances 0.000 description 1
- 238000005453 pelletization Methods 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
- 230000035699 permeability Effects 0.000 description 1
- 238000003825 pressing Methods 0.000 description 1
- 239000004449 solid propellant Substances 0.000 description 1
- 239000000126 substance Substances 0.000 description 1
- 230000001360 synchronised effect Effects 0.000 description 1
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 1
- 230000002087 whitening effect Effects 0.000 description 1
- 238000004383 yellowing Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/06—Recognition of objects for industrial automation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computational Linguistics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The embodiment of the application discloses a charge level ignition effect detection method and device based on a convolutional neural network. And finally, inputting the rest part of the charge level image as a test set to obtain a final neural network model. The model can be used for identifying the excessive generation or the excessive melting of the charge level image shot in real time, meanwhile, the false detection is removed through testing by the test set, and the identification accuracy of the abnormal state is improved.
Description
Technical Field
The embodiment of the application relates to the technical field of sintered product quality detection, in particular to a charge level ignition effect detection method and device based on a convolutional neural network.
Background
The sintering process is a key link in the iron-making process, and the working principle of the sintering process is that after a proper amount of fuel, flux and a proper amount of water are added into various powdery iron-containing raw materials, the materials are subjected to mixing and pelletizing processes, a series of physical and chemical changes are generated by using sintering equipment, and finally the materials are sintered into blocks and sent to a blast furnace to execute the next process. In the modern sintering process, the ignition furnace is used for providing high-temperature ribbon-shaped flame to the surface of the mixture, so that solid fuel in the mixture is ignited and combusted, the mixture advances forwards along with the trolley, and meanwhile, the material sintering process on the trolley is carried out downwards on the premise that oxygen is provided under the air draft effect of the exhaust fan.
However, due to the influence of various factors such as moisture, material pressing, trolley fire bars and the like, the charge level after ignition of the ignition furnace or the charge level entering the injection cover is easy to have the following two abnormal working conditions:
firstly, yellowing of local areas of the charge level (also called charge level overgrowth): the reason for the abnormal working condition is that the solid carbon particles in the fuel layer of the local area are not successfully ignited in the ignition process, and under the working condition, if the ignition intensity of the area is not timely improved, all the sintered ores produced in the area are defective ores, the yield of the sintering machine is seriously reduced, and the return ores are increased;
secondly, whitening local areas of the charge level (also called charge level overfusion): the reason for the abnormal working condition is that not only the solid carbon particles in the fuel layer are ignited in a local area in the ignition process, but also a part of iron in the iron raw material is overflown and reduced under the influence of ignition, under the working condition, if the ignition strength of the area is not timely reduced, an overflown layer is formed on the charge surface of the area, the air permeability is seriously deteriorated, and at the moment, the probability of forming defective ores is remarkably increased because the mixture at the lower layer of the area is difficult to obtain air from the upper layer.
In the prior art, in order to avoid serious reduction of the yield of the sintering machine under the two abnormal working conditions, the firing state of the charge level is usually determined manually according to visual observation and experience, so as to determine whether the opening of a valve of a sub-burner in a corresponding region of an ignition furnace needs to be integrally increased or decreased or whether the opening of a main valve of a certain work region of a gas injection device needs to be decreased. The mode consumes a large amount of manpower, and the adjusting precision is low, so that the adjusting accuracy cannot be guaranteed for operators with insufficient experience.
Disclosure of Invention
The application provides a charge level ignition effect detection method and device based on a convolutional neural network, which aim to solve the problem that the existing software cannot record and alarm abnormal flows in time.
In a first aspect, the present application provides a charge level ignition effect detection method based on a convolutional neural network, including the following steps:
acquiring a charge level image in an ignition process;
carrying out image recognition on the charge level image according to the established convolutional neural network model to obtain a recognition result; the recognition result comprises one of normal, overgrowth or overfusion.
In some embodiments, if the identification result is overgrowth or overwintering, the method further comprises:
and generating an alarm signal and sending the alarm signal to the control ends of the ignition furnace and the gas injection device.
In some embodiments, the method for building the convolutional neural network model comprises:
establishing an image set; the image set comprises a preset number of overgrowth, overfusion and normal charge level images shot in different time periods;
carrying out image preprocessing on all charge level images in the image set to obtain preprocessed charge level images;
adding artificial labels to all the preprocessed charge level images; the artificial label is one of a normal label, an overgrowth label or an overwintering label;
dividing the image set into a training set and a test set; the number of the charge level images in the training set is larger than that in the testing set;
inputting the charge level images in the training set into a residual error neural network to obtain a classification result;
comparing the classification result with an artificial label, returning an error rate and a loss function operation result to a residual error neural network, and updating a network weight value and a deviation;
and when the loss function operation result is smaller than a preset value, inputting the test set data into the residual error neural network at the moment.
In some embodiments, the step of dividing the set of images into a training set and a test set comprises:
and putting 80% of the charge level images with the same type of artificial labels into a training set, and putting the rest charge level images into a testing set.
In some embodiments, the residual neural network comprises, in order, a residual layer, a fully-connected layer, and a classifier; inputting the charge level images in the training set into a residual error neural network to obtain a classification result, wherein the classification result comprises the following steps:
inputting a charge level image to a plurality of residual error layers with different dimensionalities;
in each residual error layer, the input image passes through a convolution layer and an average pooling layer in sequence, and a first result is output;
the full connection layer performs standardized calculation on the first result to obtain a second result;
and the classifier classifies the second result to obtain a classification result.
In some embodiments, the step of adding the artificial label to all the preprocessed charge level images includes:
respectively setting RGB intervals of a normal label, an overgrowth label and an overwintering label; the upper limit value of the RGB interval is R, G, B maximum values which can be obtained respectively, and the lower limit value of the RGB interval is R, G, B minimum values which can be obtained respectively;
and correspondingly adding corresponding artificial labels to the charge level images with the RGB values falling into the RGB intervals according to the RGB values of the charge level images.
In some embodiments, the updating of the network weight values and biases employs a stochastic gradient descent method.
In some embodiments, the step of performing image preprocessing on all the charge level images in the image set to obtain a preprocessed charge level image includes:
adjusting the size of the charge level image;
adjusting the brightness of the charge level image;
and carrying out smooth denoising treatment on the adjusted charge level image.
In some embodiments, the acquiring the charge level image during the ignition process includes:
and acquiring a charge level image in the ignition process at preset time intervals.
In a second aspect, the present application also provides a convolutional neural network-based charge level ignition effect detection apparatus corresponding to the above method, including:
a camera and analysis unit; the camera is used for shooting the charge level image and sending the shot charge level image to the analysis unit; the analysis unit is configured to perform the method of the first aspect according to the received charge level image.
According to the technical scheme, the method comprises the steps of firstly collecting a large number of images of the charge level shot in different time periods under the conditions of excessive generation, excessive melting and normal, carrying out image preprocessing on the collected charge level images, then adding artificial labels to the preprocessed charge level images, wherein the artificial labels comprise three categories of normal, excessive generation and excessive melting, randomly extracting more parts of the charge level images as training sets, inputting the training sets into an established residual error neural network to carry out the characteristic extraction of the excessive generation or the excessive melting, loading the continuously input charge level image characteristics into a classifier according to an extracted multi-dimensional characteristic value matrix to carry out classification to obtain classification results, and comparing the classification results with the artificial labels to return error rate and loss function operation results to the residual error neural network. And after repeated training processes for many times, stopping training when the loss function operation result is reduced to a preset lower value, and obtaining a final multi-dimensional characteristic value matrix. And finally, inputting a small part of the charge level image as a test set to obtain a final neural network model. The model can be used for identifying the excessive generation or the excessive melting of the charge level image shot in real time, meanwhile, the false detection is removed through testing by the test set, and the identification accuracy of the abnormal state is improved.
The ignition condition of the charge level of a factory is monitored through a camera on site, images are input into a network model by utilizing a pre-trained neural network model to judge the excessive generation and excessive melting condition of the charge level, and real-time effective data are provided for an operator to adjust the total valve openness of an ignition furnace and a gas injection device.
Drawings
FIG. 1 is a step diagram of a charge level ignition effect detection method based on a convolutional neural network according to the present application;
FIG. 2 is a scene diagram of a charge level image obtained in the method provided by the present application;
FIG. 3 is a schematic diagram illustrating steps of building a convolutional neural network model in the method provided in the present application;
FIG. 4 is a schematic diagram of a normal tag added in the method provided herein;
FIG. 5 is a schematic illustration of a generic tag added in the method provided herein;
FIG. 6 is a schematic illustration of an overfused label added in the method provided herein;
FIG. 7 is a diagram of a residual neural network architecture built in accordance with the present application;
fig. 8 is a step diagram of a charge level ignition effect detection method based on a convolutional neural network according to another embodiment of the present application.
Detailed Description
Referring to fig. 1, a step diagram of a charge level ignition effect detection method based on a convolutional neural network is shown in the present application;
as can be seen from fig. 1, an embodiment of the present application provides a charge level ignition effect detection method based on a convolutional neural network, including the following steps:
s100: acquiring a charge level image in an ignition process;
in the embodiment, the charge level image can be acquired in real time, and the acquired charge level image is transmitted to a field industrial personal computer or a computer of a central control room to be analyzed and processed so as to judge whether the ignition state of the current charge level is abnormal; the process of step S100 may be generally performed by using a camera, and in a working scenario shown in fig. 2, the camera may be disposed right above the sintering site trolley, and the camera vertically shoots the charge level downwards, so that charge level images in different time periods may be obtained.
Further, according to different sintering conditions, the acquisition frequency of the charge level images can be set correspondingly, for example, the time interval for acquiring the charge level images can be set, the charge level images in the ignition process can be acquired every preset time interval, and one frame of image can be extracted from continuously shot images every 1 second interval to serve as the charge level images to be analyzed and processed.
Besides, the material level image is acquired, information related to the material level image is also included, such as the material level position, the shooting time and the like, and the subsequently obtained identification result correspondingly represents the material level state at the position and the time.
S200: carrying out image recognition on the charge level image according to the established convolutional neural network model to obtain a recognition result; the recognition result comprises one of normal, overgrowth or overfusion.
In this embodiment, before performing the identification processing on the acquired charge level image, the identification processing needs to be completed by calling the packaged classification models for identifying normal, overgrowth and overwintering, specifically, a large number of collected charge level images are input into a residual neural network for training, and finally, a convolutional neural network model with a low error rate is obtained.
As can be seen from fig. 3, the step of building the convolutional neural network model includes:
s210: establishing an image set; the image set comprises a preset number of overgrowth, overfusion and normal charge level images shot in different time periods;
in this embodiment, in order to establish a convolutional neural network model, the above-mentioned device, such as a camera, needs to be used to collect images of normal, overground and overflowed material surface in different time periods in advance, where the different time periods include physical times such as different seasons, different time periods of a day, and different time periods in an ignition process; since the level of the burden surface may be different when the burden surface is overgrown or overfused, the burden surface image also comprises images of the overgrown or overfused burden surface with different levels.
In order to ensure that the accuracy of the established convolutional neural network model is higher, tens of thousands of material level images are usually required to be trained, so the preset number in the embodiment is usually set to values of thousands to tens of thousands.
S220: carrying out image preprocessing on all charge level images in the image set to obtain preprocessed charge level images;
in the embodiment, the purpose of performing preprocessing on the charge level image is to make it easier to accurately extract the overgrowth and overwintering characteristics of the charge level. Specifically, a variety of image preprocessing methods can be used, and several possible preprocessing methods are listed below:
s221: adjusting the size of the charge level image;
firstly, there may be differences in the size of the acquired material surface image, and the image may be cut into the same size, for example, 1024 × 1024, for processing by the convolutional neural network.
S222: adjusting the brightness of the charge level image;
because the collected images are in different time periods, brightness difference may be caused by the influence of external factors (such as illumination, dust, and the like), and in order to avoid the collected images from being interfered by external factors, the brightness of the images can be adjusted, specifically, the following formula can be adopted to adjust and calculate image parameters so as to achieve better effects:
g(x，y)＝a×f(x，y)+b
wherein f (x, y) represents the numerical value of RGB three channels (red, green and blue) of pixel points of x row and y column of the source image;
g (x, y) represents the numerical value of RGB three channels (red, green and blue) of pixel points of x rows and y columns of the target image;
the parameter a (a is more than 0) represents the magnification factor (generally between 0 and 3.0);
the parameter b is called a bias parameter and is used for adjusting the brightness and is set according to actual requirements.
S223: and carrying out smooth denoising treatment on the adjusted charge level image.
In a sintering site, due to poor environment, noise interference easily occurs in an acquired image, in order to eliminate or attenuate the influence of noise in the image at the time and improve visual effect, a charge level image can be processed by adopting, for example, linear filtering, an output pixel value g (i, j) of the linear filtering process is a weighted sum of input pixel values f (i + k, j + l), and a calculation formula is as follows:
where h (k, l) is the weighting coefficient of the filter.
After the denoising processing of the filter, the pixel value difference phase of the charge level image is more obvious for the charge level under the normal condition.
It should be noted that the above steps S221 to S223 are only examples of three different image preprocessing methods, and do not represent the limitation of the sequence relationship among the three images, and in practical application, one or more of the preprocessing methods may be selected to execute the processing procedure on the charge level image, or other methods may be adopted to execute the step S220.
S230: adding artificial labels to all the preprocessed charge level images; the artificial label is one of a normal label, an overgrowth label or an overwintering label;
on the basis of preprocessing the charge level image, manual labels are made for the image, which are divided into normal, overgrowth and overwintering, the labels are respectively shown in fig. 4-6, the upper and lower limits of RGB in the figures are only examples, and other values can be set.
Further, step S230 can be decomposed into:
firstly, as illustrated in fig. 4 to 6, RGB intervals of a normal tag, an overformed tag, and an overfused tag are set respectively; the upper limit value of the RGB interval is R, G, B maximum values which can be respectively obtained, and the lower limit value of the RGB interval is R, G, B minimum values which can be respectively obtained;
and then, correspondingly adding a corresponding artificial label to the charge level image with the RGB value falling into the RGB interval according to the RGB value of the charge level image. For example, the RGB value of the face image is (147,99,38), the over-label is to be added; the RGB value of the material surface image is (88,56,45), and a normal label is added; the level image RGB value is (98,180,167) and an over-melt label is added.
S240: dividing the image set into a training set and a test set; the number of the charge level images in the training set is larger than that in the testing set;
in this embodiment, the charge level images in the training set are used to establish a residual error neural network and to execute a parameter adjustment process in the residual error neural network, and after the parameter adjustment, it is found whether there is a problem in the parameter according to a result obtained by inputting the test set to the residual error neural network, so that the test set is also used to evaluate the convolutional neural network model.
Since the training set needs to perform parameter adjustment, in order to ensure more accurate parameters, the proportion of the training set to the image set should be slightly larger, and in a feasible embodiment, 80% of the images containing the labels can be used as the training set, and the remaining 20% can be used as the test set.
Accordingly, step S240 evolves to:
and putting 80% of the charge level images with the same type of artificial labels into a training set, and putting the rest charge level images into a testing set. Specifically, the burden surface images are classified according to manual labels, 80% of the burden surface images with normal labels are selected and placed in a training set, 80% of the burden surface images with green labels are selected and placed in the training set, 80% of the burden surface images with overfused labels are selected and placed in the training set, and finally the remaining 20% of the burden surface images in the three categories are placed in a testing set.
S250: inputting the charge level images in the training set into a residual error neural network to obtain a classification result;
in this embodiment, before performing training on the charge level image in the training set, a convolutional neural network for identifying charge level overgrowth or overfusion needs to be designed, specifically, as shown in fig. 7, according to an established residual neural network structure, the residual neural network sequentially includes a residual layer, a full link layer, and a classifier; the classifier is usually a softmax classifier, but may be other types of classifiers, and is not limited herein. When performing classification, step S250 may be decomposed into:
s251: inputting a charge level image to a plurality of residual error layers with different dimensionalities; the residual error layer is designed into a plurality of different dimensions, the output of the upper layer is the input of the lower layer, and the output results are superposed layer by layer to finally obtain the superposed output result;
s252: in each residual error layer, the input image passes through a convolution layer and an average pooling layer in sequence, and a first result is output; the convolution layer can be designed to be 32-dimensional in size, in the convolution layer, input data are subjected to standardization operation and regularization operation in sequence, then convolution operation of 3 x 3 is carried out with the step length being 1, and a first result is output; the average pooling layer is used for calculating an average value of the image area as a pooled value of the area, and has the primary functions of downsampling (downsampling) dimensionality reduction, redundant information removal, feature compression, network complexity simplification, calculation amount reduction, memory consumption reduction and the like. The number is reduced and the nonlinearity is realized. The perception field can be enlarged; invariance may be implemented, where invariance includes translational invariance, rotational invariance, and scale invariance.
S253: the full connection layer performs standardized calculation on the first result to obtain a second result; the fully-connected layer is the last layer of the residual error neural network, the output of the residual error layer at the bottommost layer is the input of the fully-connected layer, the fully-connected layer mainly performs standardized calculation, regularization operation and the like of data, and finally outputs a second result;
s254: and the classifier classifies the second result to obtain a classification result. Taking the softmax classifier as an example, the softmax classifier is an unusual and important function, and is widely used in a multi-classification scene in particular, the basic principle is that some inputs are mapped to real numbers between 0 and 1, and the normalized sum is 1, so the sum of the probabilities of the multi-classification is also exactly 1, and the specific classification process is not repeated in this embodiment.
S260: comparing the classification result with an artificial label, returning an error rate and a loss function operation result to a residual error neural network, and updating a network weight value and a deviation;
in the process of inputting the charge level images in the training set into the residual error neural network, image features are loaded into a classifier for classification according to an output characteristic value matrix extracted from the charge level images, so that the residual error neural network model at the moment has the function of image classification and identification.
In the optimization algorithm for solving the machine learning parameters, a Gradient Descent-based optimization algorithm (GD) is used more frequently, and a random Gradient Descent method (SGD) algorithm is based on the GD algorithm, a group of data is extracted from samples at random, after training, a group of data is updated according to the Gradient, and then a group of data is extracted and updated again. The random is that samples are randomly disturbed in each iteration process, and the disturbance is to effectively reduce the problem of parameter updating offset caused among the samples.
S270: and when the loss function operation result is smaller than a preset value, inputting the test set data into the residual error neural network at the moment, and outputting an identification result.
In this embodiment, in order to evaluate the quality of model fitting, a loss function operation result is usually used to quantify the degree of fitting, the loss function operation result is minimized, which means that the degree of fitting is optimal, and the corresponding model parameter is the optimal parameter. In linear regression, the loss function is typically the square of the difference of the sample output and the hypothesis function. For example, for m samples (xi, yi) (i ═ 1,2, …, m), using linear regression, the loss function is:
further, when the identification result obtained in the step S200 is an overgrowth or an overwintering, in a feasible embodiment shown in fig. 8, the method further includes:
s300: generating an alarm signal and sending the alarm signal to a control end of a firing furnace and a sintering machine charge level gas injection device;
in this embodiment, the alarm signal may have various forms, such as sound, image, or email, short message reminding, etc., and the generation of the alarm signal should be synchronized with the generation of the identification result to achieve real-time performance; the control end can be a control console for manually controlling the ignition furnace and the gas injection device, and can also be an automatic control system for correspondingly executing corresponding control and adjustment operations according to the alarm signal, or an HMI (human machine interface) interface guides manual operation when no automatic adjustment device exists.
According to the technical scheme, the method comprises the steps of firstly collecting a large number of charge level images shot in different time periods, conducting image preprocessing on the collected charge level images, then adding artificial labels to the preprocessed charge level images, wherein the artificial labels comprise three categories of normal, excessive generation and excessive melting, randomly extracting more charge level images as training sets, inputting the training sets into an established residual error neural network to conduct the characteristic extraction of the excessive generation or the excessive melting, loading the continuously input charge level image characteristics into a classifier according to an extracted multi-dimensional characteristic value matrix to be classified to obtain classification results, and comparing the classification results with the artificial labels to return error rate and loss function operation results to the residual error neural network. And after repeated training processes for many times, stopping training when the loss function operation result is reduced to a preset lower value, and obtaining a final multi-dimensional characteristic value matrix. And finally, inputting a small part of the charge level image as a test set to obtain a final neural network model. The model can be used for identifying the excessive generation or the excessive melting of the charge level image shot in real time, meanwhile, the false detection is removed through testing by the test set, and the identification accuracy of the abnormal state is improved.
Corresponding to the method provided above, the present application also provides an apparatus applying the method, including: a camera and analysis unit; the camera can be an industrial camera or a monitoring camera and the like, and is used for shooting a charge level image and sending the shot charge level image to the analysis unit; the analysis unit is used for executing the method according to the received charge level image.
For the above analysis unit, it should be considered as a computer readable storage medium, a computer program product and a communication device. The computer-readable storage medium comprises instructions which, when executed on a computer, cause the computer to perform the method of any of the preceding aspects or any implementation. The computer program product, when run on a computer, causes the computer to perform the method of any of the preceding aspects or any implementation. The communication device comprises a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the method of any one of the preceding aspects or any implementation when executing the program.
The analysis unit is internally provided with a GPU processor and a trained convolutional neural network for identifying the charge level ignition effect, and can identify and alarm image information acquired by a camera in real time. The GPU processor can also be arranged in an intelligent camera, and after the camera collects images, the abnormal state monitoring is carried out through a trained convolutional neural network model in the GPU, and the result is output in real time.
Other embodiments of the invention will be apparent to those skilled in the art from consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the invention following, in general, the principles of the invention and including such departures from the present disclosure as come within known or customary practice within the art to which the invention pertains. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the invention being indicated by the following claims.
It will be understood that the invention is not limited to the precise arrangements described above and shown in the drawings and that various modifications and changes may be made without departing from the scope thereof. The scope of the invention is limited only by the appended claims.
Claims (10)
1. A charge level ignition effect detection method based on a convolutional neural network is characterized by comprising the following steps:
acquiring a charge level image in an ignition process;
carrying out image recognition on the charge level image according to the established convolutional neural network model to obtain a recognition result; the recognition result comprises one of normal, overgrowth or overfusion.
2. The method for detecting the charge level ignition effect based on the convolutional neural network as claimed in claim 1, wherein if the identification result is that the charge level ignition effect is over-generated or over-melted, the method further comprises:
and generating an alarm signal and sending the alarm signal to the control ends of the ignition furnace and the gas injection device.
3. The charge level ignition effect detection method based on the convolutional neural network as claimed in claim 1, wherein the convolutional neural network model establishing method comprises:
establishing an image set; the image set comprises a preset number of overgrowth, overfusion and normal charge level images shot in different time periods;
carrying out image preprocessing on all charge level images in the image set to obtain preprocessed charge level images;
adding artificial labels to all the preprocessed charge level images; the artificial label is one of a normal label, an overgrowth label or an overwintering label;
dividing the image set into a training set and a test set; the number of the charge level images in the training set is larger than that in the testing set;
inputting the charge level images in the training set into a residual error neural network to obtain a classification result;
comparing the classification result with an artificial label, returning an error rate and a loss function operation result to a residual error neural network, and updating a network weight value and a deviation;
and when the loss function operation result is smaller than a preset value, inputting the test set data into the residual error neural network at the moment.
4. The convolutional neural network-based charge level ignition effect detection method as claimed in claim 3, wherein the step of dividing the image set into a training set and a test set comprises:
and putting 80% of the charge level images with the same type of artificial labels into a training set, and putting the rest charge level images into a testing set.
5. The charge level ignition effect detection method based on the convolutional neural network as claimed in claim 3, wherein the residual neural network comprises a residual layer, a full-link layer and a classifier in sequence; inputting the charge level images in the training set into a residual error neural network to obtain a classification result, wherein the classification result comprises the following steps:
inputting a charge level image to a plurality of residual error layers with different dimensionalities;
in each residual error layer, the input image passes through a convolution layer and an average pooling layer in sequence, and a first result is output;
the full connection layer performs standardized calculation on the first result to obtain a second result;
and the classifier classifies the second result to obtain a classification result.
6. The method for detecting the charge level ignition effect based on the convolutional neural network as claimed in claim 3, wherein the step of adding the artificial label to all the preprocessed charge level images comprises:
respectively setting RGB intervals of a normal label, an overgrowth label and an overwintering label; the upper limit value of the RGB interval is R, G, B maximum values which can be obtained respectively, and the lower limit value of the RGB interval is R, G, B minimum values which can be obtained respectively;
and correspondingly adding corresponding artificial labels to the charge level images with the RGB values falling into the RGB intervals according to the RGB values of the charge level images.
7. The convolutional neural network-based charge level ignition effect detection method as claimed in claim 3, wherein a random gradient descent method is adopted when updating the network weight value and the deviation.
8. The charge level ignition effect detection method based on the convolutional neural network as claimed in claim 3, wherein the step of performing image preprocessing on all charge level images in the image set to obtain a preprocessed charge level image comprises:
adjusting the size of the charge level image;
adjusting the brightness of the charge level image;
and carrying out smooth denoising treatment on the adjusted charge level image.
9. The method for detecting the charge level ignition effect based on the convolutional neural network as claimed in claim 1, wherein the step of obtaining the charge level image in the ignition process comprises:
and acquiring a charge level image in the ignition process at preset time intervals.
10. A charge level ignition effect detection device based on a convolutional neural network is characterized by comprising: a camera and analysis unit; the camera is used for shooting the charge level image and sending the shot charge level image to the analysis unit; the analysis unit is configured to perform the method according to any one of claims 1 to 9 on the basis of the received charge level image.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010797640.5A CN113297886A (en) | 2020-08-10 | 2020-08-10 | Material surface ignition effect detection method and device based on convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010797640.5A CN113297886A (en) | 2020-08-10 | 2020-08-10 | Material surface ignition effect detection method and device based on convolutional neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113297886A true CN113297886A (en) | 2021-08-24 |
Family
ID=77318216
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010797640.5A Pending CN113297886A (en) | 2020-08-10 | 2020-08-10 | Material surface ignition effect detection method and device based on convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113297886A (en) |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108120311A (en) * | 2017-11-23 | 2018-06-05 | 中冶长天国际工程有限责任公司 | A kind of diagnosis of sintering machine charge level and adaptive controller and its method based on video identification |
| CN108510488A (en) * | 2018-03-30 | 2018-09-07 | 安徽理工大学 | Four kinds of damage detecting methods of conveyer belt based on residual error network |
| CN109376792A (en) * | 2018-11-07 | 2019-02-22 | 河北工业大学 | Photovoltaic cell open defect classification method based on multichannel residual error neural network |
| US20190145859A1 (en) * | 2017-11-14 | 2019-05-16 | Tula Technology, Inc. | Machine learning for misfire detection in a dynamic firing level modulation controlled engine of a vehicle |
| EP3582142A1 (en) * | 2018-06-15 | 2019-12-18 | Université de Liège | Image classification using neural networks |
| CN110728654A (en) * | 2019-09-06 | 2020-01-24 | 台州学院 | Automatic pipeline detection and classification method based on deep residual error neural network |
-
2020
- 2020-08-10 CN CN202010797640.5A patent/CN113297886A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20190145859A1 (en) * | 2017-11-14 | 2019-05-16 | Tula Technology, Inc. | Machine learning for misfire detection in a dynamic firing level modulation controlled engine of a vehicle |
| CN108120311A (en) * | 2017-11-23 | 2018-06-05 | 中冶长天国际工程有限责任公司 | A kind of diagnosis of sintering machine charge level and adaptive controller and its method based on video identification |
| CN108510488A (en) * | 2018-03-30 | 2018-09-07 | 安徽理工大学 | Four kinds of damage detecting methods of conveyer belt based on residual error network |
| EP3582142A1 (en) * | 2018-06-15 | 2019-12-18 | Université de Liège | Image classification using neural networks |
| CN109376792A (en) * | 2018-11-07 | 2019-02-22 | 河北工业大学 | Photovoltaic cell open defect classification method based on multichannel residual error neural network |
| CN110728654A (en) * | 2019-09-06 | 2020-01-24 | 台州学院 | Automatic pipeline detection and classification method based on deep residual error neural network |
Non-Patent Citations (2)
| Title |
|---|
| 刘杰 等: "基于BP神经网络的气囊点火算法模型", 《吉林大学学报（工学版）》 * |
| 猿媛之家，辛晓婷，李华荣等: "《Python程序员面试笔试宝典》", 31 March 2020 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112036755B (en) | Supervision method and system for quality detection of building engineering | |
| CN113870260B (en) | Welding defect real-time detection method and system based on high-frequency time sequence data | |
| CN114549522A (en) | Textile quality detection method based on target detection | |
| CN110057820B (en) | Method, system and storage medium for on-line detection of chlorine-hydrogen ratio of hydrogen chloride synthesis furnace | |
| CN113297885A (en) | Belt conveyor surface state detection method and device based on convolutional neural network | |
| CN107633255A (en) | A kind of rock lithology automatic recognition classification method under deep learning pattern | |
| CN112613454A (en) | Electric power infrastructure construction site violation identification method and system | |
| CN112200011B (en) | Aeration tank state detection method, system, electronic equipment and storage medium | |
| CN110909657A (en) | Method for identifying apparent tunnel disease image | |
| US20200082297A1 (en) | Inspection apparatus and machine learning method | |
| CN115995056A (en) | Automatic bridge disease identification method based on deep learning | |
| CN115115213A (en) | Method and device for retrospective analysis of quality characteristics of steel products | |
| CN114445397A (en) | Strip steel defect detection method based on shallow neural network | |
| CN113077423B (en) | Laser selective melting pool image analysis system based on convolutional neural network | |
| CN117753729A (en) | Automatic industrial dust removal equipment and control method thereof | |
| CN114066849A (en) | Deep learning-based electrical interface defect detection method | |
| CN117670755A (en) | Detection method and device for lifting hook anti-drop device, storage medium and electronic equipment | |
| CN113297886A (en) | Material surface ignition effect detection method and device based on convolutional neural network | |
| CN110320802B (en) | Complex system signal time sequence identification method based on data visualization | |
| CN115987692A (en) | Safety protection system and method based on flow backtracking analysis | |
| CN114663365A (en) | Defect detection method, device and computer storage medium | |
| CN115273009A (en) | Road crack detection method and system based on deep learning | |
| CN113052255A (en) | Intelligent detection and positioning method for reactor | |
| CN113255204A (en) | Method for calculating and counting steel-making steel material consumption by utilizing big data | |
| Tu et al. | Automatic leaf counting using improved YOLOv3 |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20210824 |
|
| RJ01 | Rejection of invention patent application after publication |