WO2020062470A1 - 基于固态激光雷达成像对煤岩界面进行识别的装置及方法 - Google Patents

基于固态激光雷达成像对煤岩界面进行识别的装置及方法 Download PDF

Info

Publication number
WO2020062470A1
WO2020062470A1 PCT/CN2018/115023 CN2018115023W WO2020062470A1 WO 2020062470 A1 WO2020062470 A1 WO 2020062470A1 CN 2018115023 W CN2018115023 W CN 2018115023W WO 2020062470 A1 WO2020062470 A1 WO 2020062470A1
Authority
WO
WIPO (PCT)
Prior art keywords
feature map
layer
coal
rock
convolution
Prior art date
Application number
PCT/CN2018/115023
Other languages
English (en)
French (fr)
Inventor
司垒
熊祥祥
王忠宾
谭超
闫海峰
姚新港
Original Assignee
中国矿业大学
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by 中国矿业大学 filed Critical 中国矿业大学
Publication of WO2020062470A1 publication Critical patent/WO2020062470A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
    • G01S7/00—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
    • G01S7/48—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S17/00
    • G01S7/4802—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S17/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • 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
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/13—Satellite images
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
    • Y02A90/30—Assessment of water resources

Definitions

  • the invention relates to the technical field of coal rock identification, in particular to a device and method for identifying a coal rock interface based on solid-state lidar imaging.
  • Coal rock identification is to identify whether the coal rock object is a coal mine or a rock.
  • coal rock identification technology can be widely used in production processes such as drum mining, heading, top coal mining, raw coal dressing, etc., to reduce the number of workers in the mining face, reduce the labor intensity of workers, improve the working environment, It is of great significance to realize safe and efficient production of coal mines and comprehensive mechanized coal mining.
  • coal and rock identification methods such as natural ray detection, stress cutting, infrared detection, active power monitoring, vibration detection, sound detection, and dust detection.
  • these methods are not universally applicable.
  • these methods are not widely used in coal and rock identification.
  • the present invention provides a device for identifying a coal rock interface based on solid-state lidar imaging.
  • the device includes : Multiple lidar modules, signal transmission modules, data storage modules, radar imaging modules, image fusion modules, and image recognition modules;
  • lidar modules are used to transmit radar signals to the same area of coal rock mine, to obtain multiple sets of coal rock mine data information in the same area of coal rock mine, wherein the lidar module can transmit radar signals to coal rock mine , And obtain the coal rock data information based on the reflected signal reflected from the coal rock mine;
  • a signal transmission module configured to transmit the plurality of groups of coal, rock, and ore data information to a data storage module
  • Data storage module used to store multiple groups of coal, rock and ore data information transmitted by the signal transmission module
  • a radar imaging module is used to retrieve multiple sets of coal, rock, and ore data information stored by the data storage module, and image each group of coal, rock, and ore data information to obtain a coal rock texture image corresponding to each group of coal, rock and ore data information, That is, multiple coal rock texture images in the same area of a coal rock mine;
  • An image fusion module configured to fuse the multiple coal rock texture images to obtain a fused coal rock texture image
  • An image recognition module is configured to perform normalization processing on the fused coal rock texture image, and recognize the normalized image to obtain a coal rock interface recognition result.
  • Each of the lidar modules includes a radar signal transmitting unit, a radar reflected signal receiving unit, and a radar signal A / D conversion unit;
  • a radar signal transmitting unit configured to transmit a radar signal to the coal rock mine
  • a radar reflection signal receiving unit configured to receive a reflection signal reflected by the coal rock ore
  • the radar signal A / D conversion unit is configured to perform data conversion on the reflection signal to obtain coal rock data information.
  • the lidar module is a solid-state lidar.
  • Normalize the existing coal and rock texture images to build a fully convolutional neural network model Use the existing coal and rock texture images after normalization to train and test the full convolutional neural network model.
  • a fully convolutional neural network model loading the trained full convolutional neural network model into the image recognition module;
  • the image recognition module performs normalization processing on the fused coal rock texture image, and inputs the normalized processed image into a trained full convolutional neural network model, and the trained full convolutional neural network
  • the network model outputs the coal-rock interface identification results.
  • the depth of the trained full convolutional neural network model is five layers, namely the first layer, the second layer, the third layer, the fourth layer, and the fifth layer;
  • the first layer consists of a convolutional layer C1, a convolutional layer C2, and a pooling layer P1.
  • Each of the convolutional layers C1 and C2 includes 64 convolution kernels of size 3 * 3 and a ReLU activation function; convolution The layer C1 is used to input the normalized image, the pixel size of the normalized image is 320 * 320 * 1, and the normalized image passes all the volumes of the convolution layer C1.
  • the feature map A1 After processing the kernel and ReLU activation function, the feature map A1 is output, and the pixel size of the feature map A1 is 318 * 318 * 64; the convolution layer C2 is used to input the feature map A1, and the feature map A1 passes all convolutions of the convolution layer C2 After processing the kernel and the ReLU activation function, the feature map A2 is output, and the pixel size of the feature map A2 is 316 * 316 * 64; the pooling layer P1 is used to input the feature map A2, and multiple 2 * 2 are divided on the feature map A2 After taking the maximum value of each block, the feature map A3 is output, and the pixel size of the feature map A3 is 158 * 158 * 64;
  • the second layer consists of a convolutional layer C3, a convolutional layer C4, and a pooling layer P2.
  • Each of the convolutional layers C3 and C4 includes 128 convolution kernels of size 2 * 2 and a ReLU activation function; convolution Layer C3 is used to input feature map A3. After processing all convolution kernels and ReLU activation functions of convolution layer C3, feature map A3 outputs feature map A4, and the pixel size of feature map A4 is 156 * 156 * 128; the convolution layer C4 is used to input feature map A4. After processing all convolution kernels and ReLU activation functions of convolution layer C4, feature map A4 outputs feature map A5.
  • the pixel size of feature map A5 is 154 * 154 * 128; pooling layer P2 It is used to input feature map A5, and divide multiple 2 * 2 blocks on feature map A5. After taking the maximum value in each block, output feature map A6, and the pixel size of feature map A6 is 77 * 77 * 128;
  • the third layer consists of a convolutional layer C5 and a convolutional layer C6.
  • Each of the convolutional layers C5 and C6 includes 256 convolution kernels of size 3 * 3 and a ReLU activation function; the convolutional layer C5 is used for input Feature map A6, feature map A6 is processed by all convolution kernels and ReLU activation functions of convolution layer C5, and output feature map A7, the pixel size of feature map A7 is 75 * 75 * 256; convolution layer C6 is used to input features Figure A7.
  • the feature map A8 is output, and the pixel size of the feature map A8 is 73 * 73 * 256;
  • the fourth layer consists of an upsampling layer U1, a convolutional layer C7, and a convolutional layer C8.
  • the upsampling layer U1 includes 256 convolution kernels of size 2 * 2.
  • Each of the convolutional layers C7 and C8 includes 128 A convolution kernel of size 3 * 3 and a ReLU activation function; the upsampling layer U1 is used to input the feature map A8, and the feature map A8 is subjected to deconvolution processing on all the convolution kernels of the upsampling layer U1 to output the feature map A9
  • the pixel size of the feature map A9 is 146 * 146 * 256; the convolution layer C7 is used to input the feature map A9.
  • the feature map A10 is output.
  • the pixel size of feature map A10 is 144 * 144 * 128; convolution layer C8 is used to input feature map A10, and feature map A10 is processed by all convolution kernels and ReLU activation functions of convolution layer C8 to output feature map A11.
  • the pixel size of Figure A11 is 142 * 142 * 128;
  • the fifth layer consists of an upsampling layer U2, a convolutional layer C9, a convolutional layer C10, and a convolutional layer C11.
  • the upsampling layer U2 includes 128 convolution kernels of size 2 * 2, a convolutional layer C9, and a convolutional layer.
  • C10 includes 64 convolution kernels of size 3 * 3 and a ReLU activation function.
  • Convolution layer C11 includes 2 convolution kernels of size 1 * 1 and a ReLU activation function.
  • the upsampling layer U2 is used to input features.
  • Figure A11 feature map A11 after all convolution kernels of the upsampling layer U2 are subjected to deconvolution, output feature map A12, the pixel size of feature map A12 is 284 * 284 * 128; convolution layer C9 is used to input the feature map A12, feature map A12 is processed by all convolution kernels and ReLU activation functions of convolution layer C9, and output feature map A13, the pixel size of feature map A13 is 282 * 282 * 64; convolution layer C10 is used to input feature map A13 After the feature map A13 is processed by all convolution kernels and ReLU activation functions of the convolution layer C10, the feature map A14 is output, and the pixel size of the feature map A14 is 280 * 280 * 64; the convolution layer C11 is used to input the feature map A14.
  • the feature map A15 and the image of the feature map A15 are output.
  • the prime size is 280 * 280 * 2.
  • the output features in Figure 15 include the coal mine interface and the rock interface of a coal rock mine.
  • the device further includes a power supply module for supplying power to the plurality of lidar modules, the signal transmission module, the data storage module, the radar imaging module, the image fusion module, and the image recognition module.
  • the present invention provides a method for identifying a coal-rock interface based on solid-state lidar imaging.
  • the method includes:
  • lidar modules are used to transmit radar signals to the same area of the coal rock mine to obtain multiple sets of coal rock mine data information, wherein each of the lidar modules can transmit radar signals to the coal rock mine, and according to the coal rock Coal mine data information is obtained by the reflection signal reflected from the mine;
  • the fused coal-rock texture image is subjected to normalization processing, and the normalized image is identified to obtain a coal-rock interface identification result.
  • the lidar modules can transmit radar signals to coal rock mines, and obtain coal rock data information based on the reflected signals reflected by the coal rock mines, including:
  • the radar signal A / D conversion unit of the lidar module performs data conversion on the reflected signal to obtain coal rock data information.
  • the normalizing the fused coal rock texture image and identifying the normalized image include:
  • the depth of the trained full convolutional neural network model is five layers, namely the first layer, the second layer, the third layer, the fourth layer, and the fifth layer;
  • the first layer consists of a convolutional layer C1, a convolutional layer C2, and a pooling layer P1.
  • Each of the convolutional layers C1 and C2 includes 64 convolution kernels of size 3 * 3 and a ReLU activation function; convolution The layer C1 is used to input the normalized image, the pixel size of the normalized image is 320 * 320 * 1, and the normalized image passes all the volumes of the convolution layer C1.
  • the feature map A1 After processing the kernel and ReLU activation function, the feature map A1 is output, and the pixel size of the feature map A1 is 318 * 318 * 64; the convolution layer C2 is used to input the feature map A1, and the feature map A1 passes all convolutions of the convolution layer C2 After processing the kernel and the ReLU activation function, the feature map A2 is output, and the pixel size of the feature map A2 is 316 * 316 * 64; the pooling layer P1 is used to input the feature map A2, and multiple 2 * 2 are divided on the feature map A2 After taking the maximum value of each block, the feature map A3 is output, and the pixel size of the feature map A3 is 158 * 158 * 64;
  • the second layer consists of a convolutional layer C3, a convolutional layer C4, and a pooling layer P2.
  • Each of the convolutional layers C3 and C4 includes 128 convolution kernels of size 2 * 2 and a ReLU activation function; convolution Layer C3 is used to input feature map A3. After processing all convolution kernels and ReLU activation functions of convolution layer C3, feature map A3 outputs feature map A4, and the pixel size of feature map A4 is 156 * 156 * 128; the convolution layer C4 is used to input feature map A4. After processing all convolution kernels and ReLU activation functions of convolution layer C4, feature map A4 outputs feature map A5.
  • the pixel size of feature map A5 is 154 * 154 * 128; pooling layer P2 It is used to input feature map A5, and divide multiple 2 * 2 blocks on feature map A5. After taking the maximum value in each block, output feature map A6, and the pixel size of feature map A6 is 77 * 77 * 128;
  • the third layer consists of a convolutional layer C5 and a convolutional layer C6.
  • Each of the convolutional layers C5 and C6 includes 256 convolution kernels of size 3 * 3 and a ReLU activation function; the convolutional layer C5 is used for input Feature map A6, feature map A6 is processed by all convolution kernels and ReLU activation functions of convolution layer C5, and output feature map A7, the pixel size of feature map A7 is 75 * 75 * 256; convolution layer C6 is used to input features Figure A7.
  • the feature map A8 is output, and the pixel size of the feature map A8 is 73 * 73 * 256;
  • the fourth layer consists of an upsampling layer U1, a convolutional layer C7, and a convolutional layer C8.
  • the upsampling layer U1 includes 256 convolution kernels of size 2 * 2.
  • Each of the convolutional layers C7 and C8 includes 128 A convolution kernel of size 3 * 3 and a ReLU activation function; the upsampling layer U1 is used to input the feature map A8, and the feature map A8 is subjected to deconvolution processing on all the convolution kernels of the upsampling layer U1 to output the feature map A9
  • the pixel size of the feature map A9 is 146 * 146 * 256; the convolution layer C7 is used to input the feature map A9.
  • the feature map A10 is output.
  • the pixel size of feature map A10 is 144 * 144 * 128; convolution layer C8 is used to input feature map A10, and feature map A10 is processed by all convolution kernels and ReLU activation functions of convolution layer C8 to output feature map A11.
  • the pixel size of Figure A11 is 142 * 142 * 128;
  • the fifth layer consists of an upsampling layer U2, a convolutional layer C9, a convolutional layer C10, and a convolutional layer C11.
  • the upsampling layer U2 includes 128 convolution kernels of size 2 * 2, a convolutional layer C9, and a convolutional layer.
  • C10 includes 64 convolution kernels of size 3 * 3 and a ReLU activation function.
  • Convolution layer C11 includes 2 convolution kernels of size 1 * 1 and a ReLU activation function.
  • the upsampling layer U2 is used to input features.
  • Figure A11 feature map A11 after all convolution kernels of the upsampling layer U2 are subjected to deconvolution, output feature map A12, the pixel size of feature map A12 is 284 * 284 * 128; convolution layer C9 is used to input the feature map A12, feature map A12 is processed by all convolution kernels and ReLU activation functions of convolution layer C9, and output feature map A13, the pixel size of feature map A13 is 282 * 282 * 64; convolution layer C10 is used to input feature map A13 After the feature map A13 is processed by all convolution kernels and ReLU activation functions of the convolution layer C10, the feature map A14 is output, and the pixel size of the feature map A14 is 280 * 280 * 64; the convolution layer C11 is used to input the feature map A14.
  • the feature map A15 and the image of the feature map A15 are output.
  • the prime size is 280 * 280 * 2.
  • the output features in Figure 15 include the coal mine interface and the rock interface of a coal rock mine.
  • the present invention has the following beneficial effects:
  • radar signals are used to detect the coal-rock interface, the detection accuracy can reach millimeter level, the relative depth of the uneven surface of coal rock can be detected, the detection process does not rely on environmental radiation, and has strong anti-interference ability.
  • Radar signals are transmitted in the same area of a coal rock mine to form multiple coal rock texture images in the same area. By merging multiple coal rock texture images in the same area, the accuracy of coal rock mine image imaging is improved, and full convolutional nerves are used.
  • the network model recognizes the coal mine interface and the rock interface on the fused coal and rock texture image, making the recognition result more accurate; the mine in the complex environment of the present invention has strong anti-interference ability, and can accurately identify the coal and rock, and The operation process is simple, the applicability is good, and the distribution of coal mines and rocks can be identified in real time.
  • FIG. 1 is a schematic structural diagram of a device for identifying a coal-rock interface based on solid-state lidar imaging according to the present invention
  • FIG. 2 is a schematic layout diagram of a lidar module of the present invention
  • FIG. 3 is a structural diagram of a fully convolutional neural network according to the present invention.
  • FIG. 4 is a flowchart of a method for identifying a coal-rock interface based on solid-state lidar imaging according to the present invention.
  • the present invention provides a device for identifying coal rock interfaces based on solid-state lidar imaging.
  • the device includes: multiple lidar modules, signal transmission modules, data storage modules, radar imaging modules, image fusion modules, and image recognition modules;
  • lidar modules are used to transmit radar signals to the same area of the coal rock mine, to obtain multiple sets of coal rock mine data information in the same area of the coal rock mine.
  • the lidar module can transmit radar signals to the coal rock mine, And obtain the coal rock data information according to the reflection signal reflected from the coal rock mine;
  • each lidar module includes a radar signal transmitting unit, a radar reflection signal receiving unit, and a radar signal A / D conversion unit;
  • Radar signal transmitting unit for transmitting radar signals to coal and rock mines
  • Radar reflected signal receiving unit used for receiving the reflected signal reflected by coal rock and ore
  • the radar signal A / D conversion unit is used to perform data conversion on the reflected signal to obtain coal rock data information.
  • the radar signal transmitting unit of each lidar module transmits a radar signal to the same area on the surface of the coal rock mine.
  • the radar signal emitted by each lidar module will be reflected on the surface of the coal rock mine, and will penetrate the coal rock mine.
  • the radar reflection signal receiving unit of each lidar module receives the reflection signal of the radar signal emitted by its own radar signal transmitting unit, and performs data conversion to obtain a set of coal rock data information. Therefore, multiple laser radars
  • the module will get multiple groups of coal rock ore data information in the same area of coal rock ore.
  • the lidar module in the present invention may be a solid-state lidar, and a solid-state lidar with a model number of CE-30 may be used.
  • a plurality of solid-state lidars are arranged in a row in front of the surface of the coal rock mine.
  • the solid-state lidar linearly moves radar signals on the surface of coal rocks and mines.
  • multiple connected areas can be divided on the surface of coal rocks and mines, which are zone 1, zone 2, zone 3 to zone.
  • each lidar module emits radar signals to the surface of coal rocks and mines in the order of area 1, area 2, area 3 to area N, so for the same area, multiple sets of coal rock data information can be obtained, as shown in Figure 2. Only three areas of the coal rock surface are shown. In the case shown in FIG. 2, three laser radar modules are used to perform laser detection of the coal rock surface in area 2.
  • a signal transmission module for transmitting multiple sets of coal, rock, and mine data information to a data storage module
  • the signal transmission module can perform data transmission through Ethernet.
  • Data storage module used to store multiple groups of coal, rock and ore data information transmitted by the signal transmission module
  • a radar imaging module is used to retrieve multiple sets of coal, rock, and ore data information stored by the data storage module, and image each group of coal, rock, and ore data information to obtain a coal rock texture image corresponding to each group of coal, rock and ore data information, That is, multiple coal rock texture images in the same area of a coal rock mine;
  • An image fusion module is used to fuse multiple coal rock texture images to obtain a fused coal rock texture image
  • An image recognition module is used to normalize the coalstone texture image after fusion, and recognize the normalized image, and obtain the coal rock interface identification result, that is, the coal mine interface and the rock interface are identified.
  • the coal and rock distribution of the coal rock ore is obtained. Among them, multiple groups of coal rock texture images in each area of the coal rock ore are imaged, fused, and identified, and the coal rock identification results of each area of the coal rock ore are obtained.
  • a data storage module, a radar imaging module, an image fusion module, and an image recognition module may be integrated in a host computer, the host computer may be a PC computer, and a solid-state laser radar may be used as a host computer.
  • a full convolutional neural network model can be used to identify coal rock texture images, specifically:
  • the existing coal rock texture image is a previously collected coal rock texture image, and the coal mine interface and rock interface corresponding to the existing coal rock texture image are also known, and the existing coal rock texture image is normalized.
  • the normalized coal rock texture image can be divided into two parts, one part is used as training data, and the other part is used as test data.
  • the training data is used to train the full convolutional neural network model, and then the test is used.
  • the data is used to test the full convolutional neural network model. If the test results have a small error with the known coal mine interface and rock interface, the test results meet the requirements. If the test results do not meet the requirements, increase the training data for the full convolutional neural network.
  • the model is trained until the test results meet the requirements, and a fully convolutional neural network model is obtained after training.
  • the trained full convolutional neural network model Load the trained full convolutional neural network model into the image recognition module, use the image recognition module to normalize the coal rock texture image fused by the image fusion module, and input the normalized image to the training completion
  • the trained full convolutional neural network model outputs the coal-rock interface recognition results, and the coal-rock recognition results include the coal-mine interface and the rock interface.
  • the depth of the trained full convolutional neural network model is five layers, namely the first layer, the second layer, the third layer, the fourth layer, and the fifth layer, as shown in FIG. 3.
  • the structure diagram of the full convolutional neural network in the present invention is the structure diagram of the full convolutional neural network in the present invention.
  • the first layer consists of a convolutional layer C1, a convolutional layer C2, and a pooling layer P1.
  • Each of the convolutional layers C1 and C2 includes 64 convolution kernels of size 3 * 3 and a ReLU activation function;
  • convolution Layer C1 is used to input the normalized image, the pixel size of the normalized image is 320 * 320 * 1, and the normalized image has all the convolution kernels and ReLUs of the convolution layer C1
  • the feature map A1 is output, and the pixel size of the feature map A1 is 318 * 318 * 64;
  • the convolution layer C2 is used to input the feature map A1, and the feature map A1 passes all the convolution kernels of the convolution layer C2.
  • the feature map A2 After performing convolution processing with the ReLU activation function, the feature map A2 is output, and the pixel size of the feature map A2 is 316 * 316 * 64; the pooling layer P1 is used to input the feature map A2, and multiple 2 are divided on the feature map A2. * 2 block, and after taking the maximum value in each block, the feature map A3 is output, and the pixel size of the feature map A3 is 158 * 158 * 64;
  • the second layer consists of a convolutional layer C3, a convolutional layer C4, and a pooling layer P2.
  • Each of the convolutional layers C3 and C4 includes 128 convolution kernels of size 2 * 2 and a ReLU activation function; convolution Layer C3 is used to input feature map A3. After feature map A3 is convolved by all convolution kernels and ReLU activation functions of convolution layer C3, feature map A4 is output, and the pixel size of feature map A4 is 156 * 156 * 128; The convolution layer C4 is used to input the feature map A4.
  • the feature map A5 is output, and the pixel size of the feature map A5 is 154 * 154 * 128;
  • the pooling layer P2 is used to input the feature map A5, and a plurality of 2 * 2 blocks are divided on the feature map A5.
  • the feature map A6 and the feature map A6 are output.
  • the pixel size is 77 * 77 * 128;
  • the third layer consists of a convolutional layer C5 and a convolutional layer C6.
  • Each of the convolutional layers C5 and C6 includes 256 convolution kernels of size 3 * 3 and a ReLU activation function; the convolutional layer C5 is used for input Feature map A6, after feature map A6 has undergone convolution processing with all convolution kernels and ReLU activation functions of convolution layer C5, feature map A7 is output, and the pixel size of feature map A7 is 75 * 75 * 256; for convolution layer C6 After inputting the feature map A7, the feature map A7 is subjected to convolution processing by all the convolution kernels and ReLU activation functions of the convolution layer C6, and the feature map A8 is output, and the pixel size of the feature map A8 is 73 * 73 * 256;
  • the fourth layer consists of an upsampling layer U1, a convolutional layer C7, and a convolutional layer C8.
  • the upsampling layer U1 includes 256 convolution kernels of size 2 * 2.
  • Each of the convolutional layers C7 and C8 includes 128 A convolution kernel of size 3 * 3 and a ReLU activation function; the upsampling layer U1 is used to input the feature map A8, and the feature map A8 is subjected to deconvolution processing on all the convolution kernels of the upsampling layer U1 to output the feature map A9
  • the pixel size of the feature map A9 is 146 * 146 * 256; the convolution layer C7 is used to input the feature map A9, and the feature map A9 is subjected to convolution processing by all the convolution kernels and ReLU activation functions of the convolution layer C7 to output features Figure A10, the pixel size of feature map A10 is 144 * 144 * 128; convolution layer C8 is used to input feature map
  • the fifth layer consists of an upsampling layer U2, a convolutional layer C9, a convolutional layer C10, and a convolutional layer C11.
  • the upsampling layer U2 includes 128 convolution kernels of size 2 * 2, a convolutional layer C9, and a convolutional layer.
  • C10 includes 64 convolution kernels of size 3 * 3 and a ReLU activation function.
  • Convolution layer C11 includes 2 convolution kernels of size 1 * 1 and a ReLU activation function.
  • the upsampling layer U2 is used to input features.
  • Figure A11 feature map A11 after all convolution kernels of the upsampling layer U2 are subjected to deconvolution, output feature map A12, the pixel size of feature map A12 is 284 * 284 * 128; convolution layer C9 is used to input the feature map A12, feature map A12 After all convolution kernels and ReLU activation functions of convolution layer C9 are used for convolution processing, feature map A13 is output, and the pixel size of feature map A13 is 282 * 282 * 64; convolution layer C10 is used for input Feature map A13, after feature map A13 undergoes convolution processing by all convolution kernels and ReLU activation functions of convolution layer C10, feature map A14 is output, and the pixel size of feature map A14 is 280 * 280 * 64; for convolution layer C11 After inputting the feature map A14, the feature map A14 is subjected to convolution processing by all the convolution kernels and ReLU activation functions of the convolution layer C11, and then input.
  • the device in the present invention may further include a power supply module for supplying power to multiple lidar modules, signal transmission modules, data storage modules, radar imaging modules, image fusion modules, and image recognition modules.
  • a power supply module for supplying power to multiple lidar modules, signal transmission modules, data storage modules, radar imaging modules, image fusion modules, and image recognition modules.
  • the data storage module, The radar imaging module, image fusion module and image recognition module can all be integrated in the host computer.
  • the device of the present invention uses a radar signal to detect a coal-rock interface.
  • the detection accuracy can reach millimeter level, and it can detect the relative depth of the uneven surface of coal-rock.
  • the detection process does not rely on environmental radiation, has strong anti-interference ability, and uses multiple lasers.
  • the radar module transmits radar signals to the same area of a coal rock mine to form multiple coal rock texture images in the same area. By merging multiple coal rock texture images in the same area, the accuracy of coal rock mine image imaging is improved.
  • the convolutional neural network model recognizes the coal mine interface and the rock interface for the fused coal-rock texture image, making the recognition result more accurate; the device in the present invention has a strong anti-interference ability in a mine in a complex environment, and can accurately It can identify coal and rock with simple operation process and good applicability. It can identify the distribution of coal and rock in real time.
  • the present invention provides a method for identifying a coal-rock interface based on solid-state lidar imaging. As shown in FIG. 4, the method includes:
  • lidar modules are used to transmit radar signals to the same area of coal rocks and mines to obtain multiple sets of coal rocks and mines data information. Among them, each lidar module can transmit radar signals to coal rocks and mines. Coal mine data information is obtained by the reflection signal reflected from the mine;
  • the lidar module includes a radar signal transmitting unit, a radar reflection signal receiving unit, and a radar signal A / D conversion unit.
  • the lidar module can transmit a radar signal to a coal rock mine, and according to the reflection reflected by the coal rock mine The signal obtains coal rock data information, including:
  • the radar signal transmitting unit of the laser radar module transmits the radar signal to the coal rock mine
  • the radar signal A / D conversion unit of the lidar module performs data conversion on the reflected signal to obtain coal rock information.
  • the radar signal transmitting unit of each lidar module transmits a radar signal to the same area on the surface of the coal rock mine.
  • the radar signal emitted by each lidar module will be reflected on the surface of the coal rock mine, and will penetrate and reflect on the coal rock mine.
  • the radar reflected signal receiving unit of each lidar module receives the reflected signal of the radar signal emitted by its own radar signal transmitting unit and performs data conversion to obtain a set of coal rock data information. Therefore, multiple lidar modules will The data of multiple groups of coal rock ore in the same area of coal rock ore are obtained.
  • the lidar module in the present invention may be a solid-state lidar.
  • a plurality of solid-state lidars are arranged in a row in front of the surface of the coal rock mine.
  • the solid-state lidar moves linearly to the coal rock mine. Radar signals occur on the surface.
  • multiple connected areas can be divided on the surface of the coal rock ore, namely area 1, area 2, area 3 to area N.
  • Each lidar module follows area 1, Area 2, area 3, and area N transmit radar signals to the surface of the coal rock and ore in sequence, so for the same area, multiple sets of coal rock ore data can be obtained.
  • the fused coal rock texture image is subjected to normalization processing, and the normalized processed image is identified.
  • a full convolutional neural network model may be used for the normalized processed image.
  • the existing coal rock texture image is a previously collected coal rock texture image, and the coal mine interface and rock interface corresponding to the existing coal rock texture image are also known, and the existing coal rock texture image is normalized.
  • the normalized coal rock texture image can be divided into two parts, one part is used as training data, and the other part is used as test data.
  • the training data is used to train the full convolutional neural network model, and then the test is used.
  • the data is used to test the full convolutional neural network model. If the test results have small errors with the known coal mine interface and rock interface, the test results meet the requirements. If the test results do not meet the requirements, increase the training data to continue to test the full convolutional neural network.
  • the network model is trained until the test results meet the requirements, and a trained fully convolutional neural network model is obtained.
  • the depth of the trained full convolutional neural network model is five layers, namely the first layer, the second layer, the third layer, the fourth layer, and the fifth layer, as shown in FIG. 3.
  • the first layer consists of a convolutional layer C1, a convolutional layer C2, and a pooling layer P1.
  • Each of the convolutional layers C1 and C2 includes 64 convolution kernels of size 3 * 3 and a ReLU activation function;
  • convolution Layer C1 is used to input the normalized image, the pixel size of the normalized image is 320 * 320 * 1, and the normalized image has all the convolution kernels and ReLUs of the convolution layer C1
  • the feature map A1 is output, and the pixel size of the feature map A1 is 318 * 318 * 64;
  • the convolution layer C2 is used to input the feature map A1, and the feature map A1 passes all the convolution kernels of the convolution layer C2.
  • the feature map A2 After performing convolution processing with the ReLU activation function, the feature map A2 is output, and the pixel size of the feature map A2 is 316 * 316 * 64; the pooling layer P1 is used to input the feature map A2, and multiple 2 are divided on the feature map A2. * 2 block, and after taking the maximum value in each block, the feature map A3 is output, and the pixel size of the feature map A3 is 158 * 158 * 64;
  • the second layer consists of a convolutional layer C3, a convolutional layer C4, and a pooling layer P2.
  • Each of the convolutional layers C3 and C4 includes 128 convolution kernels of size 2 * 2 and a ReLU activation function; convolution Layer C3 is used to input feature map A3. After feature map A3 is convolved by all convolution kernels and ReLU activation functions of convolution layer C3, feature map A4 is output, and the pixel size of feature map A4 is 156 * 156 * 128; The convolution layer C4 is used to input the feature map A4.
  • the feature map A5 is output, and the pixel size of the feature map A5 is 154 * 154 * 128;
  • the pooling layer P2 is used to input the feature map A5, and a plurality of 2 * 2 blocks are divided on the feature map A5.
  • the feature map A6 and the feature map A6 are output.
  • the pixel size is 77 * 77 * 128;
  • the third layer consists of a convolutional layer C5 and a convolutional layer C6.
  • Each of the convolutional layers C5 and C6 includes 256 convolution kernels of size 3 * 3 and a ReLU activation function; the convolutional layer C5 is used for input Feature map A6, after feature map A6 has undergone convolution processing with all convolution kernels and ReLU activation functions of convolution layer C5, feature map A7 is output, and the pixel size of feature map A7 is 75 * 75 * 256; for convolution layer C6 After inputting the feature map A7, the feature map A7 is subjected to convolution processing by all the convolution kernels and ReLU activation functions of the convolution layer C6, and the feature map A8 is output, and the pixel size of the feature map A8 is 73 * 73 * 256;
  • the fourth layer consists of an upsampling layer U1, a convolutional layer C7, and a convolutional layer C8.
  • the upsampling layer U1 includes 256 convolution kernels of size 2 * 2.
  • Each of the convolutional layers C7 and C8 includes 128 A convolution kernel of size 3 * 3 and a ReLU activation function; the upsampling layer U1 is used to input the feature map A8, and the feature map A8 is subjected to deconvolution processing on all the convolution kernels of the upsampling layer U1 to output the feature map A9
  • the pixel size of the feature map A9 is 146 * 146 * 256; the convolution layer C7 is used to input the feature map A9, and the feature map A9 is subjected to convolution processing by all the convolution kernels and ReLU activation functions of the convolution layer C7 to output features Figure A10, the pixel size of feature map A10 is 144 * 144 * 128; convolution layer C8 is used to input feature map
  • the fifth layer consists of an upsampling layer U2, a convolutional layer C9, a convolutional layer C10, and a convolutional layer C11.
  • the upsampling layer U2 includes 128 convolution kernels of size 2 * 2, a convolutional layer C9, and a convolutional layer.
  • C10 includes 64 convolution kernels of size 3 * 3 and a ReLU activation function.
  • Convolution layer C11 includes 2 convolution kernels of size 1 * 1 and a ReLU activation function.
  • the upsampling layer U2 is used to input features.
  • Figure A11 feature map A11 after all convolution kernels of the upsampling layer U2 are subjected to deconvolution, output feature map A12, the pixel size of feature map A12 is 284 * 284 * 128; convolution layer C9 is used to input the feature map A12, feature map A12 After all convolution kernels and ReLU activation functions of convolution layer C9 are used for convolution processing, feature map A13 is output, and the pixel size of feature map A13 is 282 * 282 * 64; convolution layer C10 is used for input Feature map A13, after feature map A13 undergoes convolution processing by all convolution kernels and ReLU activation functions of convolution layer C10, feature map A14 is output, and the pixel size of feature map A14 is 280 * 280 * 64; for convolution layer C11 After inputting the feature map A14, the feature map A14 is subjected to convolution processing by all the convolution kernels and ReLU activation functions of the convolution layer C11, and then input.
  • the method of the present invention uses radar signals to detect the coal-rock interface, the detection accuracy can reach millimeter level, the relative depth of the uneven surface of coal rock can be detected, the detection process does not rely on environmental radiation, and has strong anti-interference ability. It uses multiple lasers.
  • the radar module transmits radar signals to the same area of a coal rock mine to form multiple coal rock texture images in the same area. By merging multiple coal rock texture images in the same area, the accuracy of coal rock mine image imaging is improved.
  • the convolutional neural network model recognizes the coal mine interface and the rock interface on the fused coal-rock texture image, making the recognition result more accurate; the method in the present invention has a strong anti-interference ability in a mine in a complex environment, and can accurately It can identify coal and rock with simple operation process and good applicability. It can identify the distribution of coal and rock in real time.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Remote Sensing (AREA)
  • Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Radar, Positioning & Navigation (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Astronomy & Astrophysics (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

一种基于固态激光雷达成像对煤岩界面进行识别的装置，包括多个激光雷达模块、信号传输模块、数据存储模块、雷达成像模块、图像融合模块及图像识别模块；多个激光雷达模块用于向煤岩矿的同一区域发射雷达信号，得到煤岩矿同一区域的多组煤岩矿数据信息；信号传输模块用于将多组煤岩矿数据信息传输至数据存储模块；数据存储模块用于存储多组煤岩矿数据信息；雷达成像模块用于分别对每组煤岩矿数据信息进行成像，得到多个煤岩纹理图像；图像融合模块用于对多个煤岩纹理图像进行融合，得到融合后的煤岩纹理图像；图像识别模块用于对融合后的煤岩纹理图像进行归一化处理并进行识别，得出煤岩界面识别结果。还公开了一种基于固态激光雷达成像对煤岩界面进行识别的方法。

Description

基于固态激光雷达成像对煤岩界面进行识别的装置及方法 技术领域
本发明涉及煤岩识别技术领域，特别涉及一种基于固态激光雷达成像对煤岩界面进行识别的装置及方法。
背景技术
煤岩识别即识别出煤岩对象为煤矿还是为岩石。在煤炭生产过程中,煤岩识别技术可广泛应用于滚筒采煤、掘进、放顶煤开采、原煤选研石等生产环节，对于减少采掘工作面作业人员、减轻工人劳动强度、改善作业环境、实现煤矿安全高效生产和综合机械化煤炭开采具有重要意义。
现阶段，已有多种煤岩识别方法，如自然射线探测法、应力截齿法、红外探测法、有功功率监测法、震动检测法、声音检测法、粉尘检测法等。但由于煤层地质条件复杂多变,导致以上各种方法不具备普遍适用性，同时由于工作面环境恶劣、识别实时性的原因而使得这些在煤岩识别方面应用不广泛。
发明内容
为了解决现有的煤岩识别方法适用性较差，以及识别实时性不好的问题，一方面，本发明提供了一种基于固态激光雷达成像对煤岩界面进行识别的装置，所述装置包括：多个激光雷达模块、信号传输模块、数据存储模块、雷达成像模块、图像融合模块以及图像识别模块；
多个激光雷达模块，用于向煤岩矿的同一区域发射雷达信号，得到煤岩矿同一区域的多组煤岩矿数据信息，其中，所述激光雷达模块，能向煤岩矿发射雷达信号，并根据煤岩矿反射回来的反射信号得到煤岩矿数据信息；
信号传输模块，用于将所述多组煤岩矿数据信息传输至数据存储模块；
数据存储模块，用于存储信号传输模块传输过来的多组煤岩矿数据信息；
雷达成像模块，用于调取数据存储模块存储的多组煤岩矿数据信息，并分别对每组煤岩矿数据信息进行成像，得到每组煤岩矿数据信息所对应的煤岩纹理图像，即煤岩矿同一区域的多个煤岩纹理图像；
图像融合模块，用于对所述多个煤岩纹理图像进行融合，得到一张融合后的煤岩纹理图像；
图像识别模块，用于对所述融合后的煤岩纹理图像进行归一化处理，并对归一化处理后的图像进行识别，得出煤岩界面识别结果。
每个所述激光雷达模块均包括雷达信号发射单元、雷达反射信号接收单元、雷达信号A/D 转换单元；
雷达信号发射单元，用于向所述煤岩矿发射雷达信号；
雷达反射信号接收单元，用于接收被所述煤岩矿反射回来的反射信号；
雷达信号A/D转换单元，用于对所述反射信号进行数据转换，得到煤岩矿数据信息。
所述激光雷达模块为固态激光雷达。
对已有的煤岩纹理图像进行归一化处理，构建全卷积神经网络模型，利用归一化处理后的已有煤岩纹理图像对全卷积神经网络模型进行训练和测试，得到训练完毕的全卷积神经网络模型；将所述训练完毕的全卷积神经网络模型载入所述图像识别模块；
所述图像识别模块对所述融合后的煤岩纹理图像进行归一化处理，并将归一化处理后的图像输入至训练完毕的全卷积神经网络模型内，训练完毕的全卷积神经网络模型输出煤岩界面识别结果。
所述训练完毕的全卷积神经网络模型的深度为五层，分别为第一层、第二层、第三层、第四层和第五层；
第一层由卷积层C1、卷积层C2和池化层P1组成，卷积层C1和卷积层C2均包括64个大小为3*3的卷积核和一个ReLU激活函数；卷积层C1用于输入所述归一化处理后的图像，所述归一化处理后的图像像素大小为320*320*1，所述归一化处理后的图像经过卷积层C1的所有卷积核和ReLU激活函数处理后，输出特征图A1，特征图A1的像素大小为318*318*64；卷积层C2用于输入特征图A1，特征图A1经过卷积层C2的所有卷积核和ReLU激活函数处理后，输出特征图A2，特征图A2的像素大小为316*316*64；池化层P1用于输入特征图A2，并在特征图A2上划分出多个2*2的区块，并对每个区块中取最大值后，输出特征图A3，特征图A3的像素大小为158*158*64；
第二层由卷积层C3、卷积层C4和池化层P2组成，卷积层C3和卷积层C4均包括128个大小为2*2的卷积核和一个ReLU激活函数；卷积层C3用于输入特征图A3，特征图A3经过卷积层C3的所有卷积核和ReLU激活函数处理后，输出特征图A4，特征图A4的像素大小为156*156*128；卷积层C4用于输入特征图A4，特征图A4经过卷积层C4的所有卷积核和ReLU激活函数处理后，输出特征图A5，特征图A5的像素大小为154*154*128；池化层P2用于输入特征图A5，并在特征图A5上划分出多个2*2的区块，并对每个区块中取最大值后，输出特征图A6，特征图A6的像素大小为77*77*128；
第三层由卷积层C5和卷积层C6组成，卷积层C5和卷积层C6均包括256个大小为3*3的卷积核和一个ReLU激活函数；卷积层C5用于输入特征图A6，特征图A6经过卷积层C5的 所有卷积核和ReLU激活函数处理后，输出特征图A7，特征图A7的像素大小为75*75*256；卷积层C6用于输入特征图A7，特征图A7经过卷积层C6的所有卷积核和ReLU激活函数处理后，输出特征图A8，特征图A8的像素大小为73*73*256；
第四层由上采样层U1、卷积层C7和卷积层C8组成，上采样层U1包括256个大小为2*2的卷积核，卷积层C7和卷积层C8均包括128个大小为3*3的卷积核和一个ReLU激活函数；上采样层U1用于输入特征图A8，特征图A8经过上采样层U1的所有卷积核进行反卷积处理后，输出特征图A9，特征图A9的像素大小为146*146*256；卷积层C7用于输入特征图A9，特征图A9经过卷积层C7的所有卷积核和ReLU激活函数处理后，输出特征图A10，特征图A10的像素大小为144*144*128；卷积层C8用于输入特征图A10，特征图A10经过卷积层C8的所有卷积核和ReLU激活函数处理后，输出特征图A11，特征图A11的像素大小为142*142*128；
第五层由上采样层U2、卷积层C9、卷积层C10和卷积层C11组成，上采样层U2包括128个大小为2*2的卷积核，卷积层C9和卷积层C10均包括64个大小为3*3的卷积核和一个ReLU激活函数，卷积层C11包括2个大小为1*1的卷积核和一个ReLU激活函数；上采样层U2用于输入特征图A11，特征图A11经过上采样层U2的所有卷积核进行反卷积处理后，输出特征图A12，特征图A12的像素大小为284*284*128；卷积层C9用于输入特征图A12，特征图A12经过卷积层C9的所有卷积核和ReLU激活函数处理后，输出特征图A13，特征图A13的像素大小为282*282*64；卷积层C10用于输入特征图A13，特征图A13经过卷积层C10的所有卷积核和ReLU激活函数处理后，输出特征图A14，特征图A14的像素大小为280*280*64；卷积层C11用于输入特征图A14，特征图A14经过卷积层C11的所有卷积核和ReLU激活函数处理后，输出特征图A15，特征图A15的像素大小为280*280*2，特征图15的输出特征包括煤岩矿的煤矿界面和岩石界面。
所述装置还包括供电模块，用于对所述多个激光雷达模块、所述信号传输模块、所述数据存储模块、所述雷达成像模块、所述图像融合模块以及所述图像识别模块供电。
另一方面，本发明提供了一种基于固态激光雷达成像对煤岩界面进行识别的方法，所述方法包括：
采用多个激光雷达模块向煤岩矿的同一区域发射雷达信号，得到多组煤岩矿数据信息，其中，每个所述激光雷达模块，均能向煤岩矿发射雷达信号，并根据煤岩矿反射回来的反射信号得到煤岩矿数据信息；
对所述多组煤岩矿数据信息进行存储；
调取存储的多组煤岩矿数据信息，并分别对每组煤岩矿数据信息进行成像，得到每组煤 岩矿数据信息所对应的煤岩纹理图像，即煤岩矿同一区域的多个煤岩纹理图像；
对所述多个煤岩纹理图像进行融合，得到一张融合后的煤岩纹理图像；
对所述融合后的煤岩纹理图像进行归一化处理，并对归一化处理后的图像进行识别，得出煤岩界面识别结果。
所述激光雷达模块，均能向煤岩矿发射雷达信号，并根据煤岩矿反射回来的反射信号得到煤岩矿数据信息，包括：
通过所述激光雷达模块的雷达信号发射单元，向所述煤岩矿发射雷达信号；
通过所述激光雷达模块的雷达反射信号接收单元，接收被所述煤岩矿反射回来的反射信号；
通过所述激光雷达模块的雷达信号A/D转换单元，对所述反射信号进行数据转换，得到煤岩矿数据信息。
所述对所述融合后的煤岩纹理图像进行归一化处理，并对归一化处理后的图像进行识别，包括：
对已有的煤岩纹理图像进行归一化处理，构建全卷积神经网络模型，利用归一化处理后的已有煤岩纹理图像对全卷积神经网络模型进行训练和测试，得到训练完毕的全卷积神经网络模型；
对所述融合后的煤岩纹理图像进行归一化处理，并将归一化处理后的图像输入至训练完毕的全卷积神经网络模型内，训练完毕的全卷积神经网络模型输出煤岩界面识别结果。
所述训练完毕的全卷积神经网络模型的深度为五层，分别为第一层、第二层、第三层、第四层和第五层；
第一层由卷积层C1、卷积层C2和池化层P1组成，卷积层C1和卷积层C2均包括64个大小为3*3的卷积核和一个ReLU激活函数；卷积层C1用于输入所述归一化处理后的图像，所述归一化处理后的图像像素大小为320*320*1，所述归一化处理后的图像经过卷积层C1的所有卷积核和ReLU激活函数处理后，输出特征图A1，特征图A1的像素大小为318*318*64；卷积层C2用于输入特征图A1，特征图A1经过卷积层C2的所有卷积核和ReLU激活函数处理后，输出特征图A2，特征图A2的像素大小为316*316*64；池化层P1用于输入特征图A2，并在特征图A2上划分出多个2*2的区块，并对每个区块中取最大值后，输出特征图A3，特征图A3的像素大小为158*158*64；
第二层由卷积层C3、卷积层C4和池化层P2组成，卷积层C3和卷积层C4均包括128个大小为2*2的卷积核和一个ReLU激活函数；卷积层C3用于输入特征图A3，特征图A3经过 卷积层C3的所有卷积核和ReLU激活函数处理后，输出特征图A4，特征图A4的像素大小为156*156*128；卷积层C4用于输入特征图A4，特征图A4经过卷积层C4的所有卷积核和ReLU激活函数处理后，输出特征图A5，特征图A5的像素大小为154*154*128；池化层P2用于输入特征图A5，并在特征图A5上划分出多个2*2的区块，并对每个区块中取最大值后，输出特征图A6，特征图A6的像素大小为77*77*128；
第三层由卷积层C5和卷积层C6组成，卷积层C5和卷积层C6均包括256个大小为3*3的卷积核和一个ReLU激活函数；卷积层C5用于输入特征图A6，特征图A6经过卷积层C5的所有卷积核和ReLU激活函数处理后，输出特征图A7，特征图A7的像素大小为75*75*256；卷积层C6用于输入特征图A7，特征图A7经过卷积层C6的所有卷积核和ReLU激活函数处理后，输出特征图A8，特征图A8的像素大小为73*73*256；
第四层由上采样层U1、卷积层C7和卷积层C8组成，上采样层U1包括256个大小为2*2的卷积核，卷积层C7和卷积层C8均包括128个大小为3*3的卷积核和一个ReLU激活函数；上采样层U1用于输入特征图A8，特征图A8经过上采样层U1的所有卷积核进行反卷积处理后，输出特征图A9，特征图A9的像素大小为146*146*256；卷积层C7用于输入特征图A9，特征图A9经过卷积层C7的所有卷积核和ReLU激活函数处理后，输出特征图A10，特征图A10的像素大小为144*144*128；卷积层C8用于输入特征图A10，特征图A10经过卷积层C8的所有卷积核和ReLU激活函数处理后，输出特征图A11，特征图A11的像素大小为142*142*128；
第五层由上采样层U2、卷积层C9、卷积层C10和卷积层C11组成，上采样层U2包括128个大小为2*2的卷积核，卷积层C9和卷积层C10均包括64个大小为3*3的卷积核和一个ReLU激活函数，卷积层C11包括2个大小为1*1的卷积核和一个ReLU激活函数；上采样层U2用于输入特征图A11，特征图A11经过上采样层U2的所有卷积核进行反卷积处理后，输出特征图A12，特征图A12的像素大小为284*284*128；卷积层C9用于输入特征图A12，特征图A12经过卷积层C9的所有卷积核和ReLU激活函数处理后，输出特征图A13，特征图A13的像素大小为282*282*64；卷积层C10用于输入特征图A13，特征图A13经过卷积层C10的所有卷积核和ReLU激活函数处理后，输出特征图A14，特征图A14的像素大小为280*280*64；卷积层C11用于输入特征图A14，特征图A14经过卷积层C11的所有卷积核和ReLU激活函数处理后，输出特征图A15，特征图A15的像素大小为280*280*2，特征图15的输出特征包括煤岩矿的煤矿界面和岩石界面。
通过以上技术方案，相对于现有技术，本发明具有以下有益效果：
本发明中利用雷达信号对煤岩界面进行探测，探测精度能达到毫米级别，能探测煤岩凹 凸不平表面的相对深度，探测过程不依赖环境辐射，抗干扰能力强，利用多个激光雷达模块对煤岩矿的同一区域发射雷达信号，形成同一区域的多幅煤岩纹理图像，通过对同一区域的多幅煤岩纹理图像进行融合，提高煤岩矿图像成像的精确性，利用全卷积神经网络模型，对融合后的煤岩纹理图像识别煤矿界面和岩石界面，使得识别结果更加准确；本发明在复杂环境下的矿井，有很强的抗干扰能力，可以准确的进行煤岩识别，且操作过程简单，适用性较好，能对煤矿和岩石的分布情况进行实时识别。
附图说明
下面结合附图和实施例对本发明进一步说明。
图1是本发明的基于固态激光雷达成像对煤岩界面进行识别的装置的结构示意图；
图2是本发明的激光雷达模块的布置示意图；
图3是本发明的全卷积神经网络结构图；
图4是本发明的基于固态激光雷达成像对煤岩界面进行识别的方法的流程图。
具体实施方式
现在结合附图对本发明作进一步详细的说明。这些附图均为简化的示意图，仅以示意方式说明本发明的基本结构，因此其仅显示与本发明有关的构成。
实施例1
为了解决现有的煤岩识别方法适用性较差，以及识别实时性不好的问题，如图1所示，本发明提供了一种基于固态激光雷达成像对煤岩界面进行识别的装置，该装置包括：多个激光雷达模块、信号传输模块、数据存储模块、雷达成像模块、图像融合模块以及图像识别模块；
多个激光雷达模块，用于向煤岩矿的同一区域发射雷达信号，得到煤岩矿同一区域的多组煤岩矿数据信息，其中，激光雷达模块，均能向煤岩矿发射雷达信号，并根据煤岩矿反射回来的反射信号得到煤岩矿数据信息；
在本发明中，每个激光雷达模块均包括雷达信号发射单元、雷达反射信号接收单元、雷达信号A/D转换单元；
雷达信号发射单元，用于向煤岩矿发射雷达信号；
雷达反射信号接收单元，用于接收被煤岩矿反射回来的反射信号；
雷达信号A/D转换单元，用于对反射信号进行数据转换，得到煤岩矿数据信息。
其中，每个激光雷达模块的雷达信号发射单元向煤岩矿表面的同一区域发射雷达信号，每个激光雷达模块发射的雷达信号会在煤岩矿表面发生反射，还会穿透煤岩矿并进行反射， 每个激光雷达模块的雷达反射信号接收单元，接收本身的雷达信号发射单元发射的雷达信号的反射信号，并进行数据转换，得到一组煤岩矿数据信息，因此，多个激光雷达模块会得到煤岩矿同一区域的多组煤岩矿数据信息。
其中，如图2所示，本发明中的激光雷达模块可以为固态激光雷达，可以采用型号为CE-30的固态激光雷达，多个固态激光雷达排成一排置于煤岩矿表面的前方，固态激光雷达线性移动的方式对煤岩矿表面发生雷达信号，例如，如图2所示，可以在煤岩矿表面划分多个相连的区域，分别为区域1、区域2、区域3至区域N，每个激光雷达模块均按照区域1、区域2、区域3至区域N的顺序向煤岩矿表面发射雷达信号，因此对于同一区域，均可以获得多组煤岩矿数据信息，图2中仅示出了煤岩矿表面的3个区域，图2中所示的情况为，采用3个激光雷达模块对区域2的煤岩矿表面进行激光探测。
信号传输模块，用于将多组煤岩矿数据信息传输至数据存储模块；
其中，信号传输模块可以通过以太网进行数据传输。
数据存储模块，用于存储信号传输模块传输过来的多组煤岩矿数据信息；
雷达成像模块，用于调取数据存储模块存储的多组煤岩矿数据信息，并分别对每组煤岩矿数据信息进行成像，得到每组煤岩矿数据信息所对应的煤岩纹理图像，即煤岩矿同一区域的多个煤岩纹理图像；
图像融合模块，用于对多个煤岩纹理图像进行融合，得到一张融合后的煤岩纹理图像；
图像识别模块，用于对融合后的煤岩纹理图像进行归一化处理，并对归一化处理后的图像进行识别，得出煤岩界面识别结果，即识别出煤矿界面和岩石界面，如此获得了煤岩矿的煤矿和岩石的分布，其中，对煤岩矿的每个区域的多组煤岩纹理图像均进行成像、融合和识别，得到煤岩矿各个区域的煤岩识别结果。
在本发明中，数据存储模块、雷达成像模块、图像融合模块和图像识别模块可以集成在上位机内，上位机可以为PC计算机，固态激光雷达可以作为下位机。
在本发明中，可以利用全卷积神经网络模型对煤岩纹理图像进行识别，具体地：
对已有的煤岩纹理图像进行归一化处理，构建一个全卷积神经网络模型，利用归一化处理后的已有煤岩纹理图像对全卷积神经网络模型进行训练和测试，得到训练完毕的全卷积神经网络模型；
其中，已有的煤岩纹理图像为预先采集到的煤岩纹理图像，已有的煤岩纹理图像所对应的煤矿界面和岩石界面也为已知，对已有的煤岩纹理图像归一化处理后，可以将归一化处理后的已有的煤岩纹理图像分成两个部分，一部分作为训练数据，另一部分作为测试数据，利 用训练数据对全卷积神经网络模型进行训练，然后利用测试数据对全卷积神经网络模型进行测试，若测试结果与已知的煤矿界面和岩石界面误差较小，则测试结果满足要求，若测试结果不满足要求，则增加训练数据对全卷积神经网络模型进行训练，直至测试结果满足要求，得到训练完毕的全卷积神经网络模型。
将训练完毕的全卷积神经网络模型载入图像识别模块，采用图像识别模块对图像融合模块融合后的煤岩纹理图像进行归一化处理，并将归一化处理后的图像输入至训练完毕的全卷积神经网络模型内，训练完毕的全卷积神经网络模型输出煤岩界面识别结果，煤岩识别结果包括煤矿界面和岩石界面。
具体地，在本发明中，训练完毕的全卷积神经网络模型的深度为五层，分别为第一层、第二层、第三层、第四层和第五层，如图3所示的为本发明中的全卷积神经网络结构图；
第一层由卷积层C1、卷积层C2和池化层P1组成，卷积层C1和卷积层C2均包括64个大小为3*3的卷积核和一个ReLU激活函数；卷积层C1用于输入所述归一化处理后的图像，归一化处理后的图像像素大小为320*320*1，归一化处理后的图像经过卷积层C1的所有卷积核和ReLU激活函数进行卷积处理后，输出特征图A1，特征图A1的像素大小为318*318*64；卷积层C2用于输入特征图A1，特征图A1经过卷积层C2的所有卷积核和ReLU激活函数进行卷积处理后，输出特征图A2，特征图A2的像素大小为316*316*64；池化层P1用于输入特征图A2，并在特征图A2上划分出多个2*2的区块，并对每个区块中取最大值后，输出特征图A3，特征图A3的像素大小为158*158*64；
第二层由卷积层C3、卷积层C4和池化层P2组成，卷积层C3和卷积层C4均包括128个大小为2*2的卷积核和一个ReLU激活函数；卷积层C3用于输入特征图A3，特征图A3经过卷积层C3的所有卷积核和ReLU激活函数进行卷积处理后，输出特征图A4，特征图A4的像素大小为156*156*128；卷积层C4用于输入特征图A4，特征图A4经过卷积层C4的所有卷积核和ReLU激活函数进行卷积处理后，输出特征图A5，特征图A5的像素大小为154*154*128；池化层P2用于输入特征图A5，并在特征图A5上划分出多个2*2的区块，并对每个区块中取最大值后，输出特征图A6，特征图A6的像素大小为77*77*128；
第三层由卷积层C5和卷积层C6组成，卷积层C5和卷积层C6均包括256个大小为3*3的卷积核和一个ReLU激活函数；卷积层C5用于输入特征图A6，特征图A6经过卷积层C5的所有卷积核和ReLU激活函数进行卷积处理后，输出特征图A7，特征图A7的像素大小为75*75*256；卷积层C6用于输入特征图A7，特征图A7经过卷积层C6的所有卷积核和ReLU激活函数进行卷积处理后，输出特征图A8，特征图A8的像素大小为73*73*256；
第四层由上采样层U1、卷积层C7和卷积层C8组成，上采样层U1包括256个大小为2*2的卷积核，卷积层C7和卷积层C8均包括128个大小为3*3的卷积核和一个ReLU激活函数；上采样层U1用于输入特征图A8，特征图A8经过上采样层U1的所有卷积核进行反卷积处理后，输出特征图A9，特征图A9的像素大小为146*146*256；卷积层C7用于输入特征图A9，特征图A9经过卷积层C7的所有卷积核和ReLU激活函数进行卷积处理后，输出特征图A10，特征图A10的像素大小为144*144*128；卷积层C8用于输入特征图A10，特征图A10经过卷积层C8的所有卷积核和ReLU激活函数进行卷积处理后，输出特征图A11，特征图A11的像素大小为142*142*128；
第五层由上采样层U2、卷积层C9、卷积层C10和卷积层C11组成，上采样层U2包括128个大小为2*2的卷积核，卷积层C9和卷积层C10均包括64个大小为3*3的卷积核和一个ReLU激活函数，卷积层C11包括2个大小为1*1的卷积核和一个ReLU激活函数；上采样层U2用于输入特征图A11，特征图A11经过上采样层U2的所有卷积核进行反卷积处理后，输出特征图A12，特征图A12的像素大小为284*284*128；卷积层C9用于输入特征图A12，特征图A12经过卷积层C9的所有卷积核和ReLU激活函数进行卷积处理后，输出特征图A13，特征图A13的像素大小为282*282*64；卷积层C10用于输入特征图A13，特征图A13经过卷积层C10的所有卷积核和ReLU激活函数进行卷积处理后，输出特征图A14，特征图A14的像素大小为280*280*64；卷积层C11用于输入特征图A14，特征图A14经过卷积层C11的所有卷积核和ReLU激活函数进行卷积处理后，输出特征图A15，特征图A15的像素大小为280*280*2，特征图15的输出特征包括煤岩矿的煤矿界面和岩石界面。
本发明在的装置还可以包括供电模块，用于对多个激光雷达模块、信号传输模块、数据存储模块、雷达成像模块、图像融合模块以及图像识别模块供电，在本发明中，数据存储模块、雷达成像模块、图像融合模块和图像识别模块均可以集成在上位机内。
本发明中的装置，利用雷达信号对煤岩界面进行探测，探测精度能达到毫米级别，能探测煤岩凹凸不平表面的相对深度，探测过程不依赖环境辐射，抗干扰能力强，利用多个激光雷达模块对煤岩矿的同一区域发射雷达信号，形成同一区域的多幅煤岩纹理图像，通过对同一区域的多幅煤岩纹理图像进行融合，提高煤岩矿图像成像的精确性，利用全卷积神经网络模型，对融合后的煤岩纹理图像识别煤矿界面和岩石界面，使得识别结果更加准确；本发明中的装置，在复杂环境下的矿井，有很强的抗干扰能力，可以准确的进行煤岩识别，且操作过程简单，适用性较好，能对煤矿和岩石的分布情况进行实时识别。
实施例2
本发明提供一种基于固态激光雷达成像对煤岩界面进行识别的方法，如图4所示，该方法包括：
101、采用多个激光雷达模块向煤岩矿的同一区域发射雷达信号，得到多组煤岩矿数据信息，其中，每个激光雷达模块，均能向煤岩矿发射雷达信号，并根据煤岩矿反射回来的反射信号得到煤岩矿数据信息；
其中，激光雷达模块包括雷达信号发射单元、雷达反射信号接收单元和雷达信号A/D转换单元，所述激光雷达模块，均能向煤岩矿发射雷达信号，并根据煤岩矿反射回来的反射信号得到煤岩矿数据信息，包括：
通过激光雷达模块的雷达信号发射单元，向煤岩矿发射雷达信号；
通过激光雷达模块的雷达反射信号接收单元，接收被煤岩矿反射回来的反射信号；
通过激光雷达模块的雷达信号A/D转换单元，对反射信号进行数据转换，得到煤岩矿数据信息。
每个激光雷达模块的雷达信号发射单元向煤岩矿表面的同一区域发射雷达信号，每个激光雷达模块发射的雷达信号会在煤岩矿表面发生反射，还会穿透煤岩矿并进行反射，每个激光雷达模块的雷达反射信号接收单元，接收本身的雷达信号发射单元发射的雷达信号的反射信号，并进行数据转换，得到一组煤岩矿数据信息，因此，多个激光雷达模块会得到煤岩矿同一区域的多组煤岩矿数据信息。
其中，如图2所示，本发明中的激光雷达模块可以为固态激光雷达，多个固态激光雷达排成一排置于煤岩矿表面的前方，固态激光雷达线性移动的方式对煤岩矿表面发生雷达信号，例如，如图2所示，可以在煤岩矿表面划分多个相连的区域，分别为区域1、区域2、区域3至区域N，每个激光雷达模块均按照区域1、区域2、区域3至区域N的顺序向煤岩矿表面发射雷达信号，因此对于同一区域，均可以获得多组煤岩矿数据信息。
102、对多组煤岩矿数据信息进行存储；
103、调取存储的多组煤岩矿数据信息，并分别对每组煤岩矿数据信息进行成像，得到每组煤岩矿数据信息所对应的煤岩纹理图像，即煤岩矿同一区域的多个煤岩纹理图像；
104、对多个煤岩纹理图像进行融合，得到一张融合后的煤岩纹理图像；
105、对融合后的煤岩纹理图像进行归一化处理，并对归一化处理后的图像进行识别，得出煤岩界面识别结果，即识别出煤矿界面和岩石界面，如此获得了煤岩矿的煤矿和岩石的分布，对煤岩矿的每个区域的多组煤岩纹理图像均进行成像、融合和识别，得到煤岩矿各个区域的煤岩识别结果。
其中，对融合后的煤岩纹理图像进行归一化处理，并对归一化处理后的图像进行识别，在本发明中，可以利用全卷积神经网络模型对所述归一化处理后的图像进行识别，具体地：
对已有的煤岩纹理图像进行归一化处理，构建一个全卷积神经网络模型，利用归一化处理后的已有煤岩纹理图像对全卷积神经网络模型进行训练和测试，得到训练完毕的全卷积神经网络模型；
其中，已有的煤岩纹理图像为预先采集到的煤岩纹理图像，已有的煤岩纹理图像所对应的煤矿界面和岩石界面也为已知，对已有的煤岩纹理图像归一化处理后，可以将归一化处理后的已有的煤岩纹理图像分成两个部分，一部分作为训练数据，另一部分作为测试数据，利用训练数据对全卷积神经网络模型进行训练，然后利用测试数据对全卷积神经网络模型进行测试，若测试结果与已知的煤矿界面和岩石界面误差较小，则测试结果满足要求，若测试结果不满足要求，则增加训练数据继续对全卷积神经网络模型进行训练，直至测试结果满足要求，得到训练完毕的全卷积神经网络模型。
对所述融合后的煤岩纹理图像进行归一化处理，并将归一化处理后的图像输入至训练完毕的全卷积神经网络模型内，训练完毕的全卷积神经网络模型输出煤岩界面识别结果，煤岩识别结果包括煤矿界面和岩石界面。
具体地，在本发明中，训练完毕的全卷积神经网络模型的深度为五层，分别为第一层、第二层、第三层、第四层和第五层，如图3所示的为本发明中的卷积神经网络结构图；
第一层由卷积层C1、卷积层C2和池化层P1组成，卷积层C1和卷积层C2均包括64个大小为3*3的卷积核和一个ReLU激活函数；卷积层C1用于输入所述归一化处理后的图像，归一化处理后的图像像素大小为320*320*1，归一化处理后的图像经过卷积层C1的所有卷积核和ReLU激活函数进行卷积处理后，输出特征图A1，特征图A1的像素大小为318*318*64；卷积层C2用于输入特征图A1，特征图A1经过卷积层C2的所有卷积核和ReLU激活函数进行卷积处理后，输出特征图A2，特征图A2的像素大小为316*316*64；池化层P1用于输入特征图A2，并在特征图A2上划分出多个2*2的区块，并对每个区块中取最大值后，输出特征图A3，特征图A3的像素大小为158*158*64；
第二层由卷积层C3、卷积层C4和池化层P2组成，卷积层C3和卷积层C4均包括128个大小为2*2的卷积核和一个ReLU激活函数；卷积层C3用于输入特征图A3，特征图A3经过卷积层C3的所有卷积核和ReLU激活函数进行卷积处理后，输出特征图A4，特征图A4的像素大小为156*156*128；卷积层C4用于输入特征图A4，特征图A4经过卷积层C4的所有卷积核和ReLU激活函数进行卷积处理后，输出特征图A5，特征图A5的像素大小为154*154*128； 池化层P2用于输入特征图A5，并在特征图A5上划分出多个2*2的区块，并对每个区块中取最大值后，输出特征图A6，特征图A6的像素大小为77*77*128；
第三层由卷积层C5和卷积层C6组成，卷积层C5和卷积层C6均包括256个大小为3*3的卷积核和一个ReLU激活函数；卷积层C5用于输入特征图A6，特征图A6经过卷积层C5的所有卷积核和ReLU激活函数进行卷积处理后，输出特征图A7，特征图A7的像素大小为75*75*256；卷积层C6用于输入特征图A7，特征图A7经过卷积层C6的所有卷积核和ReLU激活函数进行卷积处理后，输出特征图A8，特征图A8的像素大小为73*73*256；
第四层由上采样层U1、卷积层C7和卷积层C8组成，上采样层U1包括256个大小为2*2的卷积核，卷积层C7和卷积层C8均包括128个大小为3*3的卷积核和一个ReLU激活函数；上采样层U1用于输入特征图A8，特征图A8经过上采样层U1的所有卷积核进行反卷积处理后，输出特征图A9，特征图A9的像素大小为146*146*256；卷积层C7用于输入特征图A9，特征图A9经过卷积层C7的所有卷积核和ReLU激活函数进行卷积处理后，输出特征图A10，特征图A10的像素大小为144*144*128；卷积层C8用于输入特征图A10，特征图A10经过卷积层C8的所有卷积核和ReLU激活函数进行卷积处理后，输出特征图A11，特征图A11的像素大小为142*142*128；
第五层由上采样层U2、卷积层C9、卷积层C10和卷积层C11组成，上采样层U2包括128个大小为2*2的卷积核，卷积层C9和卷积层C10均包括64个大小为3*3的卷积核和一个ReLU激活函数，卷积层C11包括2个大小为1*1的卷积核和一个ReLU激活函数；上采样层U2用于输入特征图A11，特征图A11经过上采样层U2的所有卷积核进行反卷积处理后，输出特征图A12，特征图A12的像素大小为284*284*128；卷积层C9用于输入特征图A12，特征图A12经过卷积层C9的所有卷积核和ReLU激活函数进行卷积处理后，输出特征图A13，特征图A13的像素大小为282*282*64；卷积层C10用于输入特征图A13，特征图A13经过卷积层C10的所有卷积核和ReLU激活函数进行卷积处理后，输出特征图A14，特征图A14的像素大小为280*280*64；卷积层C11用于输入特征图A14，特征图A14经过卷积层C11的所有卷积核和ReLU激活函数进行卷积处理后，输出特征图A15，特征图A15的像素大小为280*280*2，特征图15的输出特征包括煤岩矿的煤矿界面和岩石界面。
本发明中的方法，利用雷达信号对煤岩界面进行探测，探测精度能达到毫米级别，能探测煤岩凹凸不平表面的相对深度，探测过程不依赖环境辐射，抗干扰能力强，利用多个激光雷达模块对煤岩矿的同一区域发射雷达信号，形成同一区域的多幅煤岩纹理图像，通过对同一区域的多幅煤岩纹理图像进行融合，提高煤岩矿图像成像的精确性，利用全卷积神经网络 模型，对融合后的煤岩纹理图像识别煤矿界面和岩石界面，使得识别结果更加准确；本发明中的方法，在复杂环境下的矿井，有很强的抗干扰能力，可以准确的进行煤岩识别，且操作过程简单，适用性较好，能对煤矿和岩石的分布情况进行实时识别。
以上述依据本发明的理想实施例为启示，通过上述的说明内容，相关工作人员完全可以在不偏离本项发明技术思想的范围内，进行多样的变更以及修改。本项发明的技术性范围并不局限于说明书上的内容，必须要根据权利要求范围来确定其技术性范围。

Claims (10)

  1. 基于固态激光雷达成像对煤岩界面进行识别的装置，其特征在于：所述装置包括多个激光雷达模块、信号传输模块、数据存储模块、雷达成像模块、图像融合模块以及图像识别模块；
    多个激光雷达模块，用于向煤岩矿的同一区域发射雷达信号，得到煤岩矿同一区域的多组煤岩矿数据信息，其中，所述激光雷达模块，能向煤岩矿发射雷达信号，并根据煤岩矿反射回来的反射信号得到煤岩矿数据信息；
    信号传输模块，用于将所述多组煤岩矿数据信息传输至数据存储模块；
    数据存储模块，用于存储信号传输模块传输过来的多组煤岩矿数据信息；
    雷达成像模块，用于调取数据存储模块存储的多组煤岩矿数据信息，并分别对每组煤岩矿数据信息进行成像，得到每组煤岩矿数据信息所对应的煤岩纹理图像，即煤岩矿同一区域的多个煤岩纹理图像；
    图像融合模块，用于对所述多个煤岩纹理图像进行融合，得到一张融合后的煤岩纹理图像；
    图像识别模块，用于对所述融合后的煤岩纹理图像进行归一化处理，并对归一化处理后的图像进行识别，得出煤岩界面识别结果。
  2. 根据权利要求1所述的基于固态激光雷达成像对煤岩界面进行识别的装置，其特征在于：每个所述激光雷达模块均包括雷达信号发射单元、雷达反射信号接收单元、雷达信号A/D转换单元；
    雷达信号发射单元，用于向所述煤岩矿发射雷达信号；
    雷达反射信号接收单元，用于接收被所述煤岩矿反射回来的反射信号；
    雷达信号A/D转换单元，用于对所述反射信号进行数据转换，得到煤岩矿数据信息。
  3. 根据权利要求1所述的基于固态激光雷达成像对煤岩界面进行识别的装置，其特征在于：所述激光雷达模块为固态激光雷达。
  4. 根据权利要求1所述的基于固态激光雷达成像对煤岩界面进行识别的装置，其特征在于：对已有的煤岩纹理图像进行归一化处理，构建全卷积神经网络模型，利用归一化处理后的已有煤岩纹理图像对全卷积神经网络模型进行训练和测试，得到训练完毕的全卷积神经网络模型；将所述训练完毕的全卷积神经网络模型载入所述图像识别模块；
    所述图像识别模块对所述融合后的煤岩纹理图像进行归一化处理，并将归一化处理后的图像输入至训练完毕的全卷积神经网络模型内，训练完毕的全卷积神经网络模型输出煤岩界面识别结果。
  5. 根据权利要求4所述的基于固态激光雷达成像对煤岩界面进行识别的装置，其特征在于：所述训练完毕的全卷积神经网络模型的深度为五层，分别为第一层、第二层、第三层、第四层和第五层；
    第一层由卷积层C1、卷积层C2和池化层P1组成，卷积层C1和卷积层C2均包括64个大小为3*3的卷积核和一个ReLU激活函数；卷积层C1用于输入所述归一化处理后的图像，所述归一化处理后的图像像素大小为320*320*1，所述归一化处理后的图像经过卷积层C1的所有卷积核和ReLU激活函数处理后，输出特征图A1，特征图A1的像素大小为318*318*64；卷积层C2用于输入特征图A1，特征图A1经过卷积层C2的所有卷积核和ReLU激活函数处理后，输出特征图A2，特征图A2的像素大小为316*316*64；池化层P1用于输入特征图A2，并在特征图A2上划分出多个2*2的区块，并对每个区块中取最大值后，输出特征图A3，特征图A3的像素大小为158*158*64；
    第二层由卷积层C3、卷积层C4和池化层P2组成，卷积层C3和卷积层C4均包括128个大小为2*2的卷积核和一个ReLU激活函数；卷积层C3用于输入特征图A3，特征图A3经过卷积层C3的所有卷积核和ReLU激活函数处理后，输出特征图A4，特征图A4的像素大小为156*156*128；卷积层C4用于输入特征图A4，特征图A4经过卷积层C4的所有卷积核和ReLU激活函数处理后，输出特征图A5，特征图A5的像素大小为154*154*128；池化层P2用于输入特征图A5，并在特征图A5上划分出多个2*2的区块，并对每个区块中取最大值后，输出特征图A6，特征图A6的像素大小为77*77*128；
    第三层由卷积层C5和卷积层C6组成，卷积层C5和卷积层C6均包括256个大小为3*3的卷积核和一个ReLU激活函数；卷积层C5用于输入特征图A6，特征图A6经过卷积层C5的所有卷积核和ReLU激活函数处理后，输出特征图A7，特征图A7的像素大小为75*75*256；卷积层C6用于输入特征图A7，特征图A7经过卷积层C6的所有卷积核和ReLU激活函数处理后，输出特征图A8，特征图A8的像素大小为73*73*256；
    第四层由上采样层U1、卷积层C7和卷积层C8组成，上采样层U1包括256个大小为2*2的卷积核，卷积层C7和卷积层C8均包括128个大小为3*3的卷积核和一个ReLU激活函数；上采样层U1用于输入特征图A8，特征图A8经过上采样层U1的所有卷积核进行反卷积处理后，输出特征图A9，特征图A9的像素大小为146*146*256；卷积层C7用于输入特征图A9，特征图A9经过卷积层C7的所有卷积核和ReLU激活函数处理后，输出特征图A10，特征图A10的像素大小为144*144*128；卷积层C8用于输入特征图A10，特征图A10经过卷积层C8的所有卷积核和ReLU激活函数处理后，输出特征图A11，特征图A11的像素大小为142*142*128；
    第五层由上采样层U2、卷积层C9、卷积层C10和卷积层C11组成，上采样层U2包括128个大小为2*2的卷积核，卷积层C9和卷积层C10均包括64个大小为3*3的卷积核和一个ReLU激活函数，卷积层C11包括2个大小为1*1的卷积核和一个ReLU激活函数；上采样层U2用于输入特征图A11，特征图A11经过上采样层U2的所有卷积核进行反卷积处理后，输出特征图A12，特征图A12的像素大小为284*284*128；卷积层C9用于输入特征图A12，特征图A12经过卷积层C9的所有卷积核和ReLU激活函数处理后，输出特征图A13，特征图A13的像素大小为282*282*64；卷积层C10用于输入特征图A13，特征图A13经过卷积层C10的所有卷积核和ReLU激活函数处理后，输出特征图A14，特征图A14的像素大小为280*280*64；卷积层C11用于输入特征图A14，特征图A14经过卷积层C11的所有卷积核和ReLU激活函数处理后，输出特征图A15，特征图A15的像素大小为280*280*2，特征图15的输出特征包括煤岩矿的煤矿界面和岩石界面。
  6. 根据权利要求1所述的基于固态激光雷达成像对煤岩界面进行识别的装置，其特征在于：所述装置还包括供电模块，用于对所述多个激光雷达模块、所述信号传输模块、所述数据存储模块、所述雷达成像模块、所述图像融合模块以及所述图像识别模块供电。
  7. 基于固态激光雷达成像对煤岩界面进行识别的方法，其特征在于：所述方法包括：
    采用多个激光雷达模块向煤岩矿的同一区域发射雷达信号，得到多组煤岩矿数据信息，其中，每个所述激光雷达模块，均能向煤岩矿发射雷达信号，并根据煤岩矿反射回来的反射信号得到煤岩矿数据信息；
    对所述多组煤岩矿数据信息进行存储；
    调取存储的多组煤岩矿数据信息，并分别对每组煤岩矿数据信息进行成像，得到每组煤岩矿数据信息所对应的煤岩纹理图像，即煤岩矿同一区域的多个煤岩纹理图像；
    对所述多个煤岩纹理图像进行融合，得到一张融合后的煤岩纹理图像；
    对所述融合后的煤岩纹理图像进行归一化处理，并对归一化处理后的图像进行识别，得出煤岩界面识别结果。
  8. 根据权利要求7所述的基于固态激光雷达成像对煤岩界面进行识别的方法，其特征在于：所述激光雷达模块，均能向煤岩矿发射雷达信号，并根据煤岩矿反射回来的反射信号得到煤岩矿数据信息，包括：
    通过所述激光雷达模块的雷达信号发射单元，向所述煤岩矿发射雷达信号；
    通过所述激光雷达模块的雷达反射信号接收单元，接收被所述煤岩矿反射回来的反射信号；
    通过所述激光雷达模块的雷达信号A/D转换单元，对所述反射信号进行数据转换，得到煤岩矿数据信息。
  9. 根据权利要求7所述的基于固态激光雷达成像对煤岩界面进行识别的方法，其特征在于：所述对所述融合后的煤岩纹理图像进行归一化处理，并对归一化处理后的图像进行识别，包括：
    对已有的煤岩纹理图像进行归一化处理，构建全卷积神经网络模型，利用归一化处理后的已有煤岩纹理图像对全卷积神经网络模型进行训练和测试，得到训练完毕的全卷积神经网络模型；
    对所述融合后的煤岩纹理图像进行归一化处理，并将归一化处理后的图像输入至训练完毕的全卷积神经网络模型内，训练完毕的全卷积神经网络模型输出煤岩界面识别结果。
  10. 根据权利要求9所述的基于固态激光雷达成像对煤岩界面进行识别的方法，其特征在于：所述训练完毕的全卷积神经网络模型的深度为五层，分别为第一层、第二层、第三层、第四层和第五层；
    第一层由卷积层C1、卷积层C2和池化层P1组成，卷积层C1和卷积层C2均包括64个大小为3*3的卷积核和一个ReLU激活函数；卷积层C1用于输入所述归一化处理后的图像，所述归一化处理后的图像像素大小为320*320*1，所述归一化处理后的图像经过卷积层C1的所有卷积核和ReLU激活函数处理后，输出特征图A1，特征图A1的像素大小为318*318*64；卷积层C2用于输入特征图A1，特征图A1经过卷积层C2的所有卷积核和ReLU激活函数处理后，输出特征图A2，特征图A2的像素大小为316*316*64；池化层P1用于输入特征图A2，并在特征图A2上划分出多个2*2的区块，并对每个区块中取最大值后，输出特征图A3，特征图A3的像素大小为158*158*64；
    第二层由卷积层C3、卷积层C4和池化层P2组成，卷积层C3和卷积层C4均包括128个大小为2*2的卷积核和一个ReLU激活函数；卷积层C3用于输入特征图A3，特征图A3经过卷积层C3的所有卷积核和ReLU激活函数处理后，输出特征图A4，特征图A4的像素大小为156*156*128；卷积层C4用于输入特征图A4，特征图A4经过卷积层C4的所有卷积核和ReLU激活函数处理后，输出特征图A5，特征图A5的像素大小为154*154*128；池化层P2用于输入特征图A5，并在特征图A5上划分出多个2*2的区块，并对每个区块中取最大值后，输出特征图A6，特征图A6的像素大小为77*77*128；
    第三层由卷积层C5和卷积层C6组成，卷积层C5和卷积层C6均包括256个大小为3*3的卷积核和一个ReLU激活函数；卷积层C5用于输入特征图A6，特征图A6经过卷积层C5的 所有卷积核和ReLU激活函数处理后，输出特征图A7，特征图A7的像素大小为75*75*256；卷积层C6用于输入特征图A7，特征图A7经过卷积层C6的所有卷积核和ReLU激活函数处理后，输出特征图A8，特征图A8的像素大小为73*73*256；
    第四层由上采样层U1、卷积层C7和卷积层C8组成，上采样层U1包括256个大小为2*2的卷积核，卷积层C7和卷积层C8均包括128个大小为3*3的卷积核和一个ReLU激活函数；上采样层U1用于输入特征图A8，特征图A8经过上采样层U1的所有卷积核进行反卷积处理后，输出特征图A9，特征图A9的像素大小为146*146*256；卷积层C7用于输入特征图A9，特征图A9经过卷积层C7的所有卷积核和ReLU激活函数处理后，输出特征图A10，特征图A10的像素大小为144*144*128；卷积层C8用于输入特征图A10，特征图A10经过卷积层C8的所有卷积核和ReLU激活函数处理后，输出特征图A11，特征图A11的像素大小为142*142*128；
    第五层由上采样层U2、卷积层C9、卷积层C10和卷积层C11组成，上采样层U2包括128个大小为2*2的卷积核，卷积层C9和卷积层C10均包括64个大小为3*3的卷积核和一个ReLU激活函数，卷积层C11包括2个大小为1*1的卷积核和一个ReLU激活函数；上采样层U2用于输入特征图A11，特征图A11经过上采样层U2的所有卷积核进行反卷积处理后，输出特征图A12，特征图A12的像素大小为284*284*128；卷积层C9用于输入特征图A12，特征图A12经过卷积层C9的所有卷积核和ReLU激活函数处理后，输出特征图A13，特征图A13的像素大小为282*282*64；卷积层C10用于输入特征图A13，特征图A13经过卷积层C10的所有卷积核和ReLU激活函数处理后，输出特征图A14，特征图A14的像素大小为280*280*64；卷积层C11用于输入特征图A14，特征图A14经过卷积层C11的所有卷积核和ReLU激活函数处理后，输出特征图A15，特征图A15的像素大小为280*280*2，特征图15的输出特征包括煤岩矿的煤矿界面和岩石界面。
PCT/CN2018/115023 2018-09-28 2018-11-12 基于固态激光雷达成像对煤岩界面进行识别的装置及方法 WO2020062470A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201811138102.4A CN109444845B (zh) 2018-09-28 2018-09-28 基于固态激光雷达成像对煤岩界面进行识别的装置及方法
CN201811138102.4 2018-09-28

Publications (1)

Publication Number Publication Date
WO2020062470A1 true WO2020062470A1 (zh) 2020-04-02

Family

ID=65546183

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/115023 WO2020062470A1 (zh) 2018-09-28 2018-11-12 基于固态激光雷达成像对煤岩界面进行识别的装置及方法

Country Status (2)

Country Link
CN (1) CN109444845B (zh)
WO (1) WO2020062470A1 (zh)

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111337883A (zh) * 2020-04-17 2020-06-26 中国矿业大学(北京) 一种矿井煤岩界面智能探测识别系统及方法
CN111812671A (zh) * 2020-06-24 2020-10-23 北京佳力诚义科技有限公司 基于激光成像的人工智能矿石识别装置和方法
CN111931824A (zh) * 2020-07-15 2020-11-13 中煤科工集团重庆研究院有限公司 一种基于钻孔返渣图像的煤岩识别方法
CN113137230A (zh) * 2021-05-20 2021-07-20 太原理工大学 一种煤岩界面识别系统
CN113267124A (zh) * 2021-05-26 2021-08-17 济南玛恩机械电子科技有限公司 基于激光雷达的放顶煤放煤量测量系统及放煤控制方法
CN113406296A (zh) * 2021-06-24 2021-09-17 辽宁工程技术大学 一种基于深度学习的煤岩智能识别系统
CN113421222A (zh) * 2021-05-21 2021-09-21 西安科技大学 一种轻量化煤矸目标检测方法
CN114322743A (zh) * 2022-01-05 2022-04-12 瞬联软件科技（北京）有限公司 一种隧道形变实时监测系统及监测方法
CN116310843A (zh) * 2023-05-16 2023-06-23 三一重型装备有限公司 煤岩识别方法、装置、可读存储介质和掘进机

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111259095A (zh) * 2020-01-08 2020-06-09 京工博创（北京）科技有限公司 一种矿岩分界线计算方法、装置及设备
CN112001253B (zh) * 2020-07-23 2021-11-30 西安科技大学 基于改进Fast R-CNN的煤尘颗粒图像识别方法
CN111968136A (zh) * 2020-08-18 2020-11-20 华院数据技术(上海)有限公司 一种煤岩显微图像分析方法及分析系统
CN113777108B (zh) * 2021-11-10 2022-01-18 河北工业大学 双物质界面分界识别方法、设备及介质
CN116297544A (zh) * 2023-03-16 2023-06-23 南京京烁雷达科技有限公司 一种煤岩识别探地雷达目的物提取方法和装置
CN116539643B (zh) * 2023-03-16 2023-11-17 南京京烁雷达科技有限公司 一种使用雷达识别煤岩数据的方法及系统

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
SU891914A1 (ru) * 1980-03-26 1981-12-23 Научно-Производственное Объединение "Автоматгормаш Союзуглеавтоматика" Способ контрол границы "уголь-порода
US4981327A (en) * 1989-06-09 1991-01-01 Consolidation Coal Company Method and apparatus for sensing coal-rock interface
CN102496004A (zh) * 2011-11-24 2012-06-13 中国矿业大学(北京) 一种基于图像的煤岩界面识别方法与系统
CN103927514A (zh) * 2014-04-09 2014-07-16 中国矿业大学(北京) 一种基于随机局部图像特征的煤岩识别方法
CN104134074A (zh) * 2014-07-31 2014-11-05 中国矿业大学 一种基于激光扫描的煤岩识别方法
CN106845509A (zh) * 2016-10-19 2017-06-13 中国矿业大学(北京) 一种基于曲波域压缩特征的煤岩识别方法
CN107676095A (zh) * 2017-11-01 2018-02-09 天地科技股份有限公司 厚煤层放顶煤开采装置及方法

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8884806B2 (en) * 2011-10-26 2014-11-11 Raytheon Company Subterranean radar system and method
CN202383714U (zh) * 2011-11-24 2012-08-15 中国矿业大学(北京) 一种基于图像的煤岩界面识别系统
CN103207999B (zh) * 2012-11-07 2016-02-17 中国矿业大学(北京) 一种基于煤岩图像特征抽取以及分类识别的煤岩分界方法和系统
CN103472447B (zh) * 2013-09-13 2015-06-10 北京科技大学 一种基于溜槽位置判断的多点雷达协同成像装置及方法
CN107272017A (zh) * 2017-06-29 2017-10-20 深圳市速腾聚创科技有限公司 多激光雷达系统及其控制方法
CN107728143B (zh) * 2017-09-18 2021-01-19 西安电子科技大学 基于一维卷积神经网络的雷达高分辨距离像目标识别方法
CN107886121A (zh) * 2017-11-03 2018-04-06 北京清瑞维航技术发展有限公司 基于多波段雷达的目标识别方法、装置及系统
CN108564108A (zh) * 2018-03-21 2018-09-21 天津市协力自动化工程有限公司 煤炭的识别方法及装置
CN108519812B (zh) * 2018-03-21 2020-09-25 电子科技大学 一种基于卷积神经网络的三维微多普勒手势识别方法

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
SU891914A1 (ru) * 1980-03-26 1981-12-23 Научно-Производственное Объединение "Автоматгормаш Союзуглеавтоматика" Способ контрол границы "уголь-порода
US4981327A (en) * 1989-06-09 1991-01-01 Consolidation Coal Company Method and apparatus for sensing coal-rock interface
CN102496004A (zh) * 2011-11-24 2012-06-13 中国矿业大学(北京) 一种基于图像的煤岩界面识别方法与系统
CN103927514A (zh) * 2014-04-09 2014-07-16 中国矿业大学(北京) 一种基于随机局部图像特征的煤岩识别方法
CN104134074A (zh) * 2014-07-31 2014-11-05 中国矿业大学 一种基于激光扫描的煤岩识别方法
CN106845509A (zh) * 2016-10-19 2017-06-13 中国矿业大学(北京) 一种基于曲波域压缩特征的煤岩识别方法
CN107676095A (zh) * 2017-11-01 2018-02-09 天地科技股份有限公司 厚煤层放顶煤开采装置及方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
LI LIANG ET AL.: "Coal-rock interface detection method using ground penetrating radar and its experiment", INDUSTRY AND MINE AUTOMATION, vol. 41, no. 9, 30 September 2015 (2015-09-30), pages 8 - 11 *

Cited By (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111337883A (zh) * 2020-04-17 2020-06-26 中国矿业大学(北京) 一种矿井煤岩界面智能探测识别系统及方法
CN111812671A (zh) * 2020-06-24 2020-10-23 北京佳力诚义科技有限公司 基于激光成像的人工智能矿石识别装置和方法
CN111931824A (zh) * 2020-07-15 2020-11-13 中煤科工集团重庆研究院有限公司 一种基于钻孔返渣图像的煤岩识别方法
CN111931824B (zh) * 2020-07-15 2024-05-28 中煤科工集团重庆研究院有限公司 一种基于钻孔返渣图像的煤岩识别方法
CN113137230A (zh) * 2021-05-20 2021-07-20 太原理工大学 一种煤岩界面识别系统
CN113137230B (zh) * 2021-05-20 2023-08-22 太原理工大学 一种煤岩界面识别系统
CN113421222A (zh) * 2021-05-21 2021-09-21 西安科技大学 一种轻量化煤矸目标检测方法
CN113421222B (zh) * 2021-05-21 2023-06-23 西安科技大学 一种轻量化煤矸目标检测方法
CN113267124A (zh) * 2021-05-26 2021-08-17 济南玛恩机械电子科技有限公司 基于激光雷达的放顶煤放煤量测量系统及放煤控制方法
CN113406296A (zh) * 2021-06-24 2021-09-17 辽宁工程技术大学 一种基于深度学习的煤岩智能识别系统
CN114322743A (zh) * 2022-01-05 2022-04-12 瞬联软件科技（北京）有限公司 一种隧道形变实时监测系统及监测方法
CN114322743B (zh) * 2022-01-05 2024-04-12 瞬联软件科技（北京）有限公司 一种隧道形变实时监测系统及监测方法
CN116310843A (zh) * 2023-05-16 2023-06-23 三一重型装备有限公司 煤岩识别方法、装置、可读存储介质和掘进机
CN116310843B (zh) * 2023-05-16 2023-07-21 三一重型装备有限公司 煤岩识别方法、装置、可读存储介质和掘进机

Also Published As

Publication number Publication date
CN109444845A (zh) 2019-03-08
CN109444845B (zh) 2023-05-23

Similar Documents

Publication Publication Date Title
WO2020062470A1 (zh) 基于固态激光雷达成像对煤岩界面进行识别的装置及方法
CN110135468A (zh) 一种煤矸石的识别方法
US20240144467A1 (en) Hot spot defect detecting method and hot spot defect detecting system
US20190206071A1 (en) Method and apparatus for recovering point cloud data
US20190188541A1 (en) Joint 3d object detection and orientation estimation via multimodal fusion
TWI557650B (zh) 用於識別材料中之缺陷之系統及方法
CN107990876A (zh) 基于无人飞行器的地下矿山采空区快速扫描装置及方法
CN103424409B (zh) 一种基于dsp的视觉检测系统
CN108267746A (zh) 激光雷达系统、激光雷达点云数据的处理方法、可读介质
CN100555179C (zh) 一种基于cmos数字图像传感器定位装置及定位方法
CN101557452A (zh) 用于三维矿物学建模的扫描系统
US20210062653A1 (en) Method and device for acquiring three-dimensional coordinates of ore based on mining process
Wisultschew et al. 3D-LIDAR based object detection and tracking on the edge of IoT for railway level crossing
CN111651622B (zh) 一种基于三维成像的建筑用铝模板自动分类系统及其方法
CN102928441B (zh) 一种扫描式自动光学检测系统
JP2019056668A (ja) 情報処理装置、情報処理システム、情報処理方法、及び情報処理プログラム
CN112801061A (zh) 一种体态识别方法及系统
US20140180743A1 (en) Workflow monitoring and analysis system and method thereof
CN110136186A (zh) 一种用于移动机器人目标测距的检测目标匹配方法
Brandmeier et al. Lithological classification using multi-sensor data and convolutional neural networks
Xing et al. Coal and gangue identification method based on the intensity image of lidar and DenseNet
WO2022113954A1 (ja) 情報処理装置、情報処理方法、プログラム
CN104878679A (zh) 一种车载车辙实时检测系统及检测方法
WO2020215252A1 (zh) 测距装置点云滤噪的方法、测距装置和移动平台
CN114092411A (zh) 一种高效快速的双目3d点云焊点缺陷检测方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 18935744

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 18935744

Country of ref document: EP

Kind code of ref document: A1