CN111046908A

CN111046908A - Emulsion explosive package fault real-time monitoring model based on convolutional neural network

Info

Publication number: CN111046908A
Application number: CN201911071337.0A
Authority: CN
Inventors: 王越胜; 钱卓涛
Original assignee: Hangzhou Dianzi University
Current assignee: Hangzhou Dianzi University
Priority date: 2019-11-05
Filing date: 2019-11-05
Publication date: 2020-04-21

Abstract

The invention discloses an emulsion explosive package fault real-time monitoring model based on a convolutional neural network. The constructed convolutional neural network is trained by using a training set picture, then a GUI is developed by using PyQt5 based on Anaconda + Python3+ Opencv, real-time detection is performed by using a trained neural network model, and the detection can be turned on and off and a recognition result can be obtained in real time only by simple mouse clicking on a user interface. Experiments show that the method has high identification accuracy and high real-time property. The requirement of actual emulsion explosive package detection of the production line is met. The patent also provides complete code and designs the package of the complete code into a GUI interface, so that a user can directly use and read the code and improve programs on the basis of the code.

Description

Emulsion explosive package fault real-time monitoring model based on convolutional neural network

Technical Field

The invention belongs to the technical field of fault detection, and relates to an emulsion explosive package fault real-time monitoring model based on a convolutional neural network.

Background

Currently, the monitoring of the package failure of the emulsion explosive commonly used in the industry mainly depends on the following two modes:

firstly, manual observation; the manual observation method mainly depends on the workers in the explosive factory to observe whether the packaged emulsion explosive is abnormal in real time, the conveyor belt transportation is stopped if the abnormality is found, the reason for the abnormality of the emulsion explosive package is found, then the conveyor belt is opened, and the production line continues to run. The manual observation and detection has strong subjectivity, and the long-time work easily causes the fatigue of workers, the condition of overlooking or mislooking often easily occurs, the operation efficiency of the production line of a explosive plant is influenced, and irreparable consequences such as explosion and the like can be caused in serious cases.

Secondly, the propagation process obtains an external matrix of the emulsion explosive package in the research on the online detection technology and application of the industrial explosive packaging process, and distinguishes and detects the emulsion explosive package faults according to the length, width and length-width ratio of the external matrix; the appearance of the emulsion explosive cartridge is cylindrical, the industrial field cartridge is packaged in orange, the transmission belt is green, the emulsion explosive is bright brown, and the cartridge, the emulsion explosive and the transmission belt have obvious brightness difference in images acquired by a black-and-white area array camera. When the packaging machine is used for packaging cartridges, due to external factors such as the opening and stopping processes of the machine, too much or too little cartridges can be filled, if too much, bags can be expanded, and if too much, the cartridges can be packaged and damaged, and emulsion explosives can leak, so that the production line is polluted. And the situation that the two ends of the medicated roll are leaked due to the fact that the two ends of the medicated roll are not firmly buckled when the two ends of the medicated roll are buckled is also possible. The general emulsion explosive package detection comprises the conditions of normal explosive sticks, side-end explosive-leaking explosive sticks, port explosive-leaking explosive sticks, package sunken explosive sticks, leaked emulsion explosives and the like. The external shape information of the drug roll can be known through the image, such as information of length, width, perimeter, area, space coordinates and the like. Through analysis and summary of a large number of pictures on site, the defective cartridge can be identified through the contour size.

In summary, the problems of the prior art are as follows:

(1) the manual observation and detection has strong subjectivity, and the long-time work easily causes the fatigue of workers, the condition of overlooking or mislooking often easily occurs, the operation efficiency of the production line of a explosive plant is influenced, and irreparable consequences such as explosion and the like can be caused in serious cases.

(2) The method for distinguishing and detecting the package faults of the emulsified explosive mainly uses a BP neural network according to the length, width and length-width ratio of an external matrix, but the distinguishing accuracy rate is not high and is generally between 80 and 90 percent; and the algorithm has slow running time, is not suitable for processing videos shot in real time, only can continuously shoot pictures by using an industrial camera and then transmit the pictures to a BP neural network for identification, and has very low real-time property.

(3) Without complete program code, the user cannot directly extract the test and obtain the final recognition result.

(4) At present, a complete independent user interface is not developed, and the identification of the shot emulsion explosive package photos is researched, and the practicability of the emulsion explosive package photos is not researched.

The difficulty in solving the technical problems is as follows:

1: the accuracy rate of identifying the package faults of the emulsion explosive is not high, and particularly, the difference between the expansion bag and the normal external rectangle of the emulsion explosive is not much, so that the expansion bag and the normal situation can not be accurately distinguished.

2: in order to achieve higher real-time monitoring difficulty and achieve higher real-time performance under certain hardware conditions, the convolutional neural network needs to be simplified and improved, so that the operation rate is improved while the accuracy is ensured. The white blood cells cannot be classified and counted, and the counting can only be used for learning research and cannot be used in an actual production line.

3: the problem that in the prior art, only the emulsion explosive on a production line can be photographed firstly and then identified, and the specific position of the fault packaged emulsion explosive which cannot be accurately positioned and identified is caused by time delay is urgently solved.

Disclosure of Invention

Aiming at the problems in the prior art, the invention provides a convolutional neural network-based emulsion explosive package fault real-time monitoring model.

The technical scheme adopted by the invention for solving the technical problems is as follows:

step 1, reading picture pixels, wherein the read picture is a 28 × 1 gray picture, the first layer is a convolution layer, the convolution kernel size is 5 × 5, the depth is 6, and the step size is 1;

the convolution layer is directly connected with the input layer, and the type of pictures which can be read by the improved convolution neural network model is a gray-scale picture with the size of 28 × 1. The first convolutional kernel size is 5x5, depth is 6, all 0 padding is not used, and step size is 1. Since all 0 padding is not used, the node matrix for the convolutional layer output is 28-5+ 1-24 with a depth of 6; namely 24 x 6 node matrix i.

Step 2, the second layer adopts a maximum pooling layer with the filter of 2x2, the length, the width and the step length of the filter are both 2, so the size of the node matrix II of the output of the second layer is 12 x 6;

the pooling layer is directly connected with the output of the previous layer, the output of the previous layer after convolution processing is a 24X 6-3456 node matrix, and the pooling layer changes the 24X 6 node matrix I of the output of the previous layer into 12X 6 node matrix II by using a filter with length, width and motion step length of 2 and outputs the node matrix I to the next layer. The pooling layer employs a maximum pooling method, i.e., sequentially retaining the maximum of four pixels in 2x2, and the pooling operation can reduce the amount of information carried by the matrix to

Step 3, the third layer is a convolution layer, wherein the size of a convolution kernel is 5 × 5, the depth is 16, the step size is 1, all 0 filling is not adopted in the same convolution layer, and the node matrix II of the previous layer 12 × 6 can be changed into a node matrix III of 8 × 16;

the convolutional layer is connected to the second (pooling layer) output, the convolutional layer convolution kernel size is 5x5, depth is 16, no all 0 padding is used, step size is 1. Since all 0 padding is not used, the output of the convolutional layer has a node matrix iii of 12-5+ 1-8 and a depth of 16, i.e., 8-16.

Step 4, adopting a maximum pooling layer with a filter of 2x2 in the fourth layer, wherein the length, the width and the step length of the filter are both 2, so that the size of an output matrix IV of the maximum pooling layer is 4 x 16;

the maximum pooling layer is directly connected with the output of the last convolution layer, and the output of the last convolution layer is 8 × 16 ═ 1024 node matrix III; the maximum pooling layer uses a filter with length and width and motion step of 2 to change the node matrix III of 8 x 16 outputted from the previous layer into a node matrix IV of 4 x 16 outputted to the next layer. The pooling layer adopts a maximum pooling method, namely, the maximum of four pixels in 2x2 is reserved in sequence, and the information amount carried by the matrix can be reduced to the value

Step 5, the fifth layer is an improved full-connection layer, the full-connection layer and a LeNet-5 convolutional neural network model are greatly changed, and the original three full-connection layers are simplified into one full-connection layer;

the fully-connected layer is connected with the output of the last maximum pooling layer, the output of the last maximum pooling layer after the maximum pooling process is a node matrix IV with 4 x 16 x 256, and the fully-connected layer has the function of drawing the nodes of the node matrix IV into a vector, so that the fully-connected layer receives a 256-dimensional input vector. Each node of the full-connection layer is connected with nodes of the upper layer and the lower layer.

Step 6, the sixth layer is an output layer, and the output layer has 4 nodes because the defects of the emulsified explosive packages distinguished by the experiment are 4 in total;

the output layer is connected with the last full connection layer, the full connection layer has 256 output nodes, and each node is connected with four nodes of the output layer. Each node of the output layer represents a defect in the package of 4 experimentally differentiated emulsion explosives.

Step 7, mode testing

Selecting 4 types of packages of normally packaged emulsion explosive, excessive (bag expansion), unsaturated (head distortion) and leaked explosive, identifying and classifying by utilizing a trained improved LeNet-5 convolutional neural network, and carrying out 10 tests in the experiment, wherein the classification accuracy and the loss function value of each test on the 100 pictures are shown in the following table 1:

TABLE 1 recognition accuracy statistics table

And (4) analyzing results: from the recognition accuracy of 10 experiments, the recognition accuracy of 10 experiments reaches over 90%, the recognition rate of individual experiments reaches over 97%, loss function values of LeNet-5 convolutional neural network models are recorded after training is completed in each training, the loss function values of 10 experiments are very small, and the convolutional neural network models are very consistent with the method for processing and recognizing emulsion explosive package faults, the number of training set pictures of the convolutional neural network is 1100 in total, the number of model training times is 1000 at present, and the accuracy can be further improved if the number of the training set pictures is increased continuously and the proper number of model training times is increased.

In summary, the advantages and positive effects of the invention are:

the method adopts the improved LeNet-5-based convolutional neural network for identification and classification, and the convolutional neural network has good image identification and classification effects; in the experiment, 4 situations of normally packaged emulsion explosive, excessive explosive loading (bag expansion), unsaturated explosive loading (head bending) and explosive leakage can be clearly distinguished, and the identification image is visual and clear and has high accuracy. The accuracy of identifying the faults of the emulsion explosive package is improved, the real-time performance of identification is improved, and the fault types of whether the emulsion explosive package is normal or not and abnormal packages can be identified when the emulsion explosive on the conveyer belt enters the shooting range of a camera; the staff can be according to the trouble packing that detects out in the video, and the concrete position of this emulsion explosive is fixed a position to quick accuracy. Therefore, the fault detection also has practical significance, and can be applied to the actual explosive production instead of only staying at the research level.

Drawings

Fig. 1 is a flow chart provided by an embodiment of the present invention.

FIG. 2 is a training process of the improved LeNet-5 convolutional neural network model of the present invention

Fig. 3 is a schematic diagram of an improved LeNet-5 convolutional neural network model provided by an embodiment of the present invention.

Fig. 4 is a grayscale image of an emulsion explosive provided by an embodiment of the invention.

FIG. 5 is a binary image of an emulsion explosive provided by an embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail with reference to the following embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

The constructed convolutional neural network is trained by using a training set picture, then a GUI is developed by using PyQt5 based on Anaconda + Python3+ Opencv, real-time detection is performed by using a trained neural network model, and the detection can be turned on and off and a recognition result can be obtained in real time only by simple mouse clicking on a user interface. Experiments show that the method has high identification accuracy and high real-time property. The requirement of actual emulsion explosive package detection of the production line is met. The patent also provides complete code and designs the package of the complete code into a GUI interface, so that a user can directly use and read the code and improve programs on the basis of the code.

The invention will be further described in detail with reference to the accompanying drawings.

As shown in figure 1, the invention provides a convolutional neural network-based emulsion explosive package fault real-time monitoring model. The specific implementation comprises the following steps:

s101: reading picture pixels, wherein the readable picture is a 28 × 1 gray picture, the first layer is a convolution layer, the convolution kernel size is 5 × 5, the depth is 6, and the step size is 1;

the specific codes are as follows:

conv1_weights＝tf.get_variable("conv1_weights",[5,5,1,32],initializer＝tf.truncated_normal_initializer(stddev＝0.1))

conv1_biases＝tf.get_variable("conv1_biases",[32],initializer＝tf.constant_initializer(0.0))

conv1＝tf.nn.conv2d(x_image,conv1_weights,strides＝[1,1,1,1],padding＝'VALID')

relu1＝tf.nn.relu(tf.nn.bias_add(conv1,conv1_biases))

s102: the second layer adopts a maximum pooling layer with the filter of 2x2, the length and the width of the filter and the step size of the filter are both 2, so the output matrix size of the layer is 12 x 6;

the specific codes are as follows:

pool1＝tf.nn.max_pool(relu1,ksize＝[1,2,2,1],strides＝[1,2,2,1],padding＝'VALID')

s103: the third layer is a convolution layer, wherein the size of a convolution kernel is 5 × 5, the depth is 16, the step size is 1, the same layer is not filled with all 0, and the node matrix of the previous layer 12 × 6 can be changed into a node matrix of 8 × 16;

the specific codes are as follows:

conv2_weights＝tf.get_variable("conv2_weights",[5,5,32,64],initializer＝tf.truncated_normal_initializer(stddev＝0.1))

conv2_biases＝tf.get_variable("conv2_biases",[64],initializer＝tf.constant_initializer(0.0))

conv2＝tf.nn.conv2d(pool1,conv2_weights,strides＝[1,1,1,1],padding＝'VALID')

relu2＝tf.nn.relu(tf.nn.bias_add(conv2,conv2_biases))

s104: the fourth layer adopts a maximum pooling layer with the filter of 2x2, the length and the width of the filter and the step size of the filter are both 2, so the output matrix size of the layer is 4 x 16;

the specific codes are as follows:

pool2＝tf.nn.max_pool(relu2,ksize＝[1,2,2,1],strides＝[1,2,2,1],padding＝'VALID')

s105: the fifth layer is a full connection layer, which is greatly changed with a LeNet-5 convolutional neural network model, and the original three full connection layers are simplified into one full connection layer;

the specific codes are as follows:

fc1 weight tg _ variable ("fc1 weight", [4 × 16,256], initializer tf. truncated _ normal _ initializer (stddev 0.1)) #7 × 7 64 3136 change the output of the previous layer into a feature vector

fc1_baises＝tf.get_variable("fc1_baises",[1024],initializer＝tf.constant_initializer(0.1))

pool2_vector＝tf.reshape(pool2,[-1,4*4*16])

fc1＝tf.nn.relu(tf.matmul(pool2_vector,fc1_weights)+fc1_baises)

S106: the sixth layer is an output layer, and the defects of the emulsified explosive packages distinguished in the experiment are 4 in total, so that the output layer has 4 nodes;

the specific codes are as follows:

y_conv＝tf.nn.softmax(fc1)

the invention is further described with reference to specific examples.

The emulsion explosive package fault real-time monitoring technology and system based on the improved LeNet-5 convolutional neural network model provided by the embodiment of the invention comprise:

reading picture pixels, wherein the readable picture is a 28 × 1 gray picture, the first layer is a convolution layer, the convolution kernel size is 5 × 5, the depth is 6, and the step size is 1;

the second layer adopts a maximum pooling layer with the filter of 2x2, the length and the width of the filter and the step size of the filter are both 2, so the output matrix size of the layer is 12 x 6;

the third layer is a convolution layer, wherein the size of a convolution kernel is 5 × 5, the depth is 16, the step size is 1, the same layer is not filled with all 0, and the node matrix of the previous layer 12 × 6 can be changed into a node matrix of 8 × 16;

the fourth layer adopts a maximum pooling layer with the filter of 2x2, the length and the width of the filter and the step size of the filter are both 2, so the output matrix size of the layer is 4 x 16;

the fifth layer is a full connection layer, which is greatly changed with a LeNet-5 convolutional neural network model, and the original three full connection layers are simplified into one full connection layer;

the sixth layer is an output layer, and the defects of the emulsified explosive packages distinguished in the experiment are 4 in total, so that the output layer has 4 nodes.

The development and implementation of the invention are based on Anaconda + Python3+ Opencv, and a GUI interface is developed by PyQt5, so that a user can easily control the operation of a detection program on the interface and display a fault detection result in real time.

The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents and improvements made within the spirit and principle of the present invention are intended to be included within the scope of the present invention.

Claims

1. the real-time monitoring model of emulsion explosive packaging fault based on convolutional neural network, is characterized in that comprising the steps:

Step 1. Read the picture pixels. The readable picture is a grayscale picture of 28*28*1. The first layer is a convolution layer, where the convolution kernel size is 5*5, the depth is 6, and the step size is 1. ;

The convolutional layer is directly connected to the input layer. The image type that can be read by this improved convolutional neural network model is a grayscale image with a size of 32*32*1; the size of the convolution kernel of the first convolutional layer is 5x5, The depth is 6, all 0 padding is not used, and the stride is 1; since all 0 padding is not used, the node matrix of the output of the convolutional layer is 28-5+1=24, and the depth is 6; that is, 24*24* 6 node matrix I;

Step 2. The second layer adopts a maximum pooling layer with a filter of 2x2. The length and width of the filter and the step size are both 2, so the size of the output node matrix II of the second layer is 12*12*6;

The pooling layer is directly connected to the output of the previous convolutional layer. The output of the previous layer is a 24*24*6=3456 node matrix after convolution processing. The pooling layer uses a length and width and a motion step size All filters are 2, and the 24*24*6 node matrix I output by the previous layer is changed into a 12*12*6 node matrix II and output to the next layer; the pooling layer adopts the maximum pooling method, that is, keep the maximum value of the four pixels in 2*2 in turn, and the pooling operation can reduce the amount of information carried by the matrix to

Step 3. The third layer is the convolution layer, where the size of the convolution kernel is 5*5, the depth is 16, and the step size is 1. The same convolution layer does not use all 0 padding, and the previous layer can be 12* The node matrix II of 12*6 is changed to the node matrix III of 8*8*16;

The convolutional layer is connected to the output of the second layer (pooling layer). The size of the convolutional kernel of the convolutional layer is 5x5, the depth is 16, the padding of all 0s is not used, and the stride is 1; since padding of all 0s is not used, Therefore, the output node matrix III of the convolution layer is 12-5+1=8, and the depth is 16, that is, the node matrix III of 8*8*16;

Step 4. The fourth layer adopts a maximum pooling layer with a filter of 2x2. The length and width of the filter and the step size are both 2, so the output matrix IV of the maximum pooling layer is 4*4*16 in size;

The maximum pooling layer is directly connected to the output of the previous convolutional layer, and the output of the previous layer is a node matrix III of 8*8*16=1024 after convolution processing; the maximum pooling layer uses a length and width and A filter with a motion step size of 2 changes the 8*8*16 node matrix III output by the previous layer into a 4*4*16 node matrix IV and outputs it to the next layer; the pooling layer uses the maximum The value pooling method, that is, the maximum value of the four pixels in 2*2 is retained in turn. After this pooling operation, the amount of information carried by the matrix can be reduced to

Step 5. The fifth layer is an improved fully connected layer, which is greatly changed from the LeNet-5 convolutional neural network model, and the original three fully connected layers are simplified into one fully connected layer;

The fully connected layer is connected to the output of the previous maximum pooling layer. The output of the previous maximum pooling layer is a 4*4*16=256 node matrix IV after the maximum pooling process, and the function of this fully connected layer is to Pull the nodes of the node matrix IV into a vector, then the fully connected layer receives a 256-dimensional input vector; each node of the fully connected layer is connected to the nodes of the previous layer and the next layer;

Step 6. The sixth layer is the output layer. Since there are 4 kinds of emulsion explosive packaging defects distinguished in the experiment, the output layer has 4 nodes; this output layer is connected to the previous fully connected layer, and the fully connected layer has a total of 256 output nodes. Each node is connected to four nodes of the output layer; each node of the output layer represents the four experimentally differentiated emulsion explosive packaging defects.