WO2020220674A1 - 一种基于次卷积超相关的垃圾识别分类处理深度学习方法 - Google Patents

一种基于次卷积超相关的垃圾识别分类处理深度学习方法 Download PDF

Info

Publication number
WO2020220674A1
WO2020220674A1 PCT/CN2019/122533 CN2019122533W WO2020220674A1 WO 2020220674 A1 WO2020220674 A1 WO 2020220674A1 CN 2019122533 W CN2019122533 W CN 2019122533W WO 2020220674 A1 WO2020220674 A1 WO 2020220674A1
Authority
WO
WIPO (PCT)
Prior art keywords
deep learning
module
garbage
correlation
sub
Prior art date
Application number
PCT/CN2019/122533
Other languages
English (en)
French (fr)
Inventor
黄骏
王洁
徐童
Original Assignee
宿迁海沁节能科技有限公司
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 宿迁海沁节能科技有限公司 filed Critical 宿迁海沁节能科技有限公司
Publication of WO2020220674A1 publication Critical patent/WO2020220674A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N33/00—Investigating or analysing materials by specific methods not covered by groups G01N1/00 - G01N31/00
    • G01N33/0004—Gaseous mixtures, e.g. polluted air
    • G01N33/0009—General constructional details of gas analysers, e.g. portable test equipment
    • G01N33/0027—General constructional details of gas analysers, e.g. portable test equipment concerning the detector
    • G01N33/0031—General constructional details of gas analysers, e.g. portable test equipment concerning the detector comprising two or more sensors, e.g. a sensor array
    • 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
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
    • G06F18/21355—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis nonlinear criteria, e.g. embedding a manifold in a Euclidean space
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks

Definitions

  • the invention belongs to the technical field of electronic classification, and in particular relates to a deep learning method for garbage identification and classification processing based on sub-convolution super correlation.
  • Artificial neural network abstracts the human brain neuron network from the perspective of information processing to establish a simple model.
  • Animal nerve cells contain both the amplitude information (real part operation) and phase information (imaginary part operation) of the stimulation signal.
  • the existing The neural network algorithm does not have any phase information.
  • the existing deep learning is mainly divided into three main schools: the deep belief neural network used by Google, the convolutional neural network used by facebook, and the posterior probabilistic neural network used by Microsoft. These three different neural networks are designed to avoid artificial nerves.
  • Network amplitude information overflows between layers, but they are all real number axes that only use electronic calculations and not imaginary number axes that use quantum calculations.
  • the real axis is a line that can record a limited amount of information, and it is very easy to overflow during operations. With the imaginary axis stimulus signal distributed on a disk surface, the space of the activity track is larger and it is not easy to overflow.
  • Coal mine plants can produce solid waste and gas waste during the production process.
  • Solid waste is generally treated by landfill, chemical conversion, incineration and other methods. Since garbage is generally not classified during recycling, severe disturbances may occur between different solids during processing. Biochemical reactions sometimes generate additional biogas due to the relatively closed processing environment. Biogas is flammable and explosive when mixed with air, which may damage the processing equipment. The main component of biogas is methane, which causes a greenhouse effect greater than carbon dioxide. Methane diffuses more easily than air, and it is difficult to concentrate treatment after diffusion. Separate collection means identifying garbage and collecting garbage with the same attributes. Gas waste generally does not need to be identified, only whether the gas is toxic or not needs to be checked.
  • the existing solid waste treatment methods may produce violent reactions between different solids, generating additional toxic gases, and causing damage to the human body and the environment.
  • the present invention provides a deep learning method for garbage identification and classification processing based on sub-convolutional super correlation.
  • a deep learning method for garbage identification and classification processing based on subconvolutional super correlation includes:
  • the first step is to input a set of sequences to the deep learning convolutional layer, and according to the generalized binomial formula, obtain the coefficients of the generalized binomial of the sequence expressed by the gamma function;
  • the second step is to perform convolution with the weight matrix of the convolution layer after the input data is subjected to sub-convolution;
  • the third step is to perform hypervariance calculation on the corresponding output data obtained
  • the fourth step is to get whether the object in the camera's field of view is garbage in the mine, and know what kind of garbage the object is;
  • the obtained information is transferred to the control module, and the control module triggers the robotic hand to reclaim the object or clean it on site to complete the entire reclaiming and processing process.
  • the first step specifically includes: the input of the deep learning convolutional layer is a set of sequences x 1 , x 2 , ... x n , and the Euclidean norm is:
  • a x When all a is greater than 0 and a ⁇ 1, x ⁇ R, a x can be defined as:
  • ⁇ 1.5 is the input of the deep learning convolutional layer for sub-convolution operation.
  • the input data is subjected to sub-convolution for operation, it is convolved with the weight matrix w of the convolutional layer.
  • Another object of the present invention is to provide a garbage recognition classification processing deep learning system based on sub-convolutional super correlation, including:
  • the control system is connected with the deep learning module, camera, radio module, gas recognition module, drive module, peripheral processing module, distance measurement module, GPS module, function: as the control center of a single deep learning robot, control the above modules to complete all functions .
  • the control system consists of CPU, RAM, ROM, and bus.
  • the deep learning module is connected with the control system and the camera, and is used to receive the data passed by the camera for deep learning, and use the learned frame to identify the object shot by the camera and determine the attributes of the object.
  • the camera is connected with the control system and the deep learning module, used to transmit road condition information that the robot has driven, and provide analysis data for the deep learning module;
  • the radio module is connected to the control system for communication with the control center and other robots; the radio module mainly uses an industrial-grade eSIM card, which is small and can communicate with the control center at a long distance.
  • the gas recognition module is connected to the control system to identify toxic gases and detect the gas in real time; the gas recognition module is composed of multiple gas sensors.
  • the drive module is connected to the control system and used to generate corresponding magnetism by the change of current strength to provide power for the robot; it is composed of an iron shell and an internal magnetic attraction.
  • Peripheral module connected to the control system, initially equipped with a robotic arm for grabbing solid waste, and advanced equipment with a multi-mode sprinkler for cleaning dust and waste on the spot;
  • Ranging module connected to the control system, used to emit laser beams to calculate the distance between the robot and the object;
  • the GPS module is connected to the control system to obtain real-time position information.
  • Another object of the present invention is to provide a robot wirelessly connected to the garbage recognition and classification processing deep learning system based on subconvolution super correlation.
  • the garbage positioning and ranging function of the present invention when there is solid garbage in the field of view of the camera, the laser emitted from the laser toward the front of the robot is used, and the photoelectric element receives the laser reflected 3 At the same time, the timer records the time between the launch and the reception of the laser beam.
  • the modified robot uses GPS positioning technology, so that the control center can get the location of each robot in real time. When encountering a task that a robot cannot complete, the control center controls the arrival of neighboring robots according to the position of each robot, and cooperates to complete it.
  • the deep learning recognition, classification and processing garbage of the present invention when an object appears in the camera field of view, the control module triggers the deep learning module to quickly identify the object and perform classification processing on it.
  • the identification of toxic gas in the present invention using a gas sensor, when a toxic gas is recognized, the control module sends a toxic gas location message to the command center to provide location information for subsequent work.
  • the present invention has the following advantages: (1) When a certain object is collected, when encountering an impassable space, the object in the collection device is taken out, and the current position is located at the same time, and the appearance is reduced. Bring the objects that were dropped on the return journey. (2) The new algorithm is used to improve the traditional deep learning, which reduces the time to recognize objects, and is more accurate in analyzing the types of objects. It will not treat garbage as flowers or flowers as garbage. (3) The change of the current strength changes the magnitude of the magnetic force and provides power for the movement of the robot. (4) The command center can control multiple robots to work at the same time, reducing the difficulty of collection.
  • the present invention is an improved deep learning method, which runs faster than general deep learning.
  • the invention integrates the deep learning recognition and classification function into the garbage robot. When the robot's camera scans garbage, it automatically triggers the deep learning recognition mode, classifies it into the corresponding category, and uses robotic hands to collect garbage or clean it on the spot.
  • the invention can be used not only in the sanitation cleaning system, but also in any dangerous situations that cannot be reached by human beings.
  • the device of the invention weight 5 ⁇ 7(tons) 1 ⁇ 2(tons) size 3 ⁇ 5(m) 1 ⁇ 3(m) Clean up the premises can can Water the flowers Can't can Collect gas Can't can Collect solids can can Collect liquid Can't can AI does not have have
  • FIG. 1 is a schematic structural diagram of a deep learning system for garbage identification and classification based on subconvolutional super correlation provided by an embodiment of the present invention
  • Control system In the figure: 1. Control system; 2. Deep learning module; 3. Camera; 4. Radio module; 5. Gas recognition module; 6. Drive module; 7. Peripheral processing module; 8. Ranging module; 9. GPS Module.
  • Fig. 2 is a schematic structural diagram of a control module provided by an embodiment of the present invention.
  • Fig. 3 is a schematic structural diagram of a driving module provided by an embodiment of the present invention.
  • Fig. 4 is a schematic structural diagram of a ranging module provided by an embodiment of the present invention.
  • FIG. 5 is a flowchart of a deep learning method for garbage identification and classification processing based on subconvolutional super correlation provided by an embodiment of the present invention.
  • the invention utilizes embedded technology, garbage positioning technology, signal processing and identification technology in the mine robot, and realizes the functions of automatic identification, classification and recycling of the garbage. At the same time, it has the function of gas recovery and cooperation with multiple robots.
  • the garbage recognition and classification processing deep learning system based on subconvolutional super correlation includes: a control system 1, a deep learning module 2, a camera 3, a radio module 4, and a gas recognition module 5.
  • Control system 1 used to connect all relevant components
  • Deep learning module 2 connected to control system 1, used to classify and process different scenarios
  • the camera 3 is connected with the deep learning module 2 to transmit road condition information that the robot has driven and provide analysis data for the deep learning module 2.
  • the radio module 4 connected with the control system 1, is used to coordinate with surrounding companions
  • the gas recognition module 5 is connected to the control system 1 and is used to recognize toxic gas and detect the gas in real time.
  • the drive module 6 is connected to the control system 1 and is used to generate corresponding magnetism by using the current strength change to provide power for the robot.
  • Peripheral processing module 7 connected with control system 1, used for grabbing and recycling or cleaning in place
  • the ranging module 8 is connected to the control system 1 and is used to emit a laser beam to calculate the distance between the robot and the object.
  • the GPS module 9 is connected to the control system 1 for obtaining real-time position information.
  • the driving module 6 of the garbage recognition and classification deep learning system based on sub-convolutional super-correlation provided by the embodiment of the present invention generates corresponding magnetism by using changes in current strength to provide power for the robot.
  • the control module 1 transmits garbage that needs to be collected
  • the wireless device emits a laser beam to calculate the distance between the robot and the object; then the manipulator extends from the back of the robot to obtain the object and put it into the collection device; or according to different types of flowers Non-destructive angle and pulsed water pressure in-situ cleaning. If it encounters objects that are difficult to collect, the command center can obtain real-time position information through the GPS module 9 and control the closer robots to complete it together.
  • the robot adopts the design of concentric circles, the periphery is the collection space, and the deformable technology is adopted, which can be freely contracted; the inside is the control module of the entire robot; the gas sensor must detect the gas in real time in order to identify the toxic gas.
  • the deep learning method for garbage identification and classification based on sub-convolutional super correlation includes the following steps:
  • S101 Input a set of sequences to the deep learning convolutional layer, and obtain the binomial coefficients of the sequence according to the binomial formula;
  • S104 Obtain whether the object in the camera's field of view is mine garbage, and know what kind of garbage the object is;
  • S105 The obtained information is transmitted to the control module, and the control module triggers the robotic hand to reclaim the object or clean it in place to complete the entire reclaiming process.
  • step S101 according to the generalized binomial formula, the coefficients of the generalized binomial of the sequence expressed by the gamma function are obtained
  • the deep learning method for garbage recognition classification processing based on subconvolutional super correlation adopts an improved convolutional neural network.
  • the identification method is:
  • the input of the deep learning convolutional layer is a set of sequences x 1 , x 2 ,...x n , the Euclidean norm is:
  • a x When all a is greater than 0 and a ⁇ 1, x ⁇ R, a x can be defined as:
  • ⁇ 1.5 is the input of the deep learning convolutional layer for sub-convolution operation.
  • the second step is to perform convolution with the weight matrix w of the convolution layer after the input data is operated by subconvolution.
  • the obtained information is passed to the control module, and the control module triggers the robotic hand to recycle or clean the object to complete the entire recycling process.
  • Table 1 proves that adding complex number operations to high-dimensional deep convolution can effectively improve the recognition accuracy.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Chemical & Material Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Health & Medical Sciences (AREA)
  • Evolutionary Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Analytical Chemistry (AREA)
  • Pathology (AREA)
  • Immunology (AREA)
  • Biochemistry (AREA)
  • Medicinal Chemistry (AREA)
  • Food Science & Technology (AREA)
  • Combustion & Propulsion (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Processing Of Solid Wastes (AREA)
  • Image Analysis (AREA)

Abstract

一种基于次卷积超相关的垃圾识别分类处理深度学习方法，首先，深度学习卷积层输入一组序列，根据广义二项式公式，得到序列的广义二项式以伽马函数表达的系数；当输入的数据用次卷积做运算后，与卷积层的权重矩阵做卷积，得到的相应输出数据，做超方差运算；之后通过输出数据判断摄像头视野中的物体是否为矿场垃圾，并且知道该物体是哪一类垃圾；最后，将所得到的信息传递给控制模块，控制模块触发机器手回收物体，或者外设模块清洗物体，完成整个回收或就地处理过程。

Description

一种基于次卷积超相关的垃圾识别分类处理深度学习方法 技术领域
本发明属于电子分类技术领域，尤其涉及一种基于次卷积超相关的垃圾识别分类处理深度学习方法。
背景技术
目前，最接近的现有技术：
随着工业生产的发展，工业废物数量日益增加。然而工业废物得到利用的只有有限的几种，例如日本等国利用了粉煤灰和煤渣，美国等国利用钢铁渣。大多数工业废物以堆存为主，部分有害固体采用焚烧、填埋、化学转化等方法进行处理。煤炭是我国重要的基础能源和重要原料，近年来我国煤炭工业得到高速发展，但其安全问题和垃圾处理问题依然处于严峻状态。煤矿里的垃圾一般燃点都很低，长时间的堆积极易自燃，并且污染空气、水和植被。各种城市规划科技园汽车保有量逐年上升，路边花卉每日被蒙上一层厚厚的汽车有毒尾气的灰尘，容易导致各种大面积传染的植物疾病，也需要每天按时清洗。人工处理以上环境污染都不是一件容易的事，工作人员伤亡事故屡见不鲜。
人工神经网络是从信息处理角度对人脑神经元网络进行抽象建立某种简单模型，动物神经细胞既包含刺激信号的幅度信息(实部运算)又包含相位信息(虚部运算)，然而现有的神经网络算法并无任何的相位信息。现有的深度学习主要分为三大流派：谷歌使用的深度置信神经网络、facebook采用的卷积神经网络、微软使用的后验概率神经网络，这三种不同的神经网络为的就是避免人工神经网络幅度信息的层间溢出，然而他们都是只使用电子计算的实数轴没有使用量子计算的虚数轴，实轴是一条线所能记录的信息量有限，在进行运算时非常容易溢出。有了虚数轴的刺激信号分布在一个圆盘面上，活动轨迹的空间更大就不容易溢出。
煤矿厂在生产过程中可以产生固体垃圾和气体垃圾，固体垃圾一般采用填埋、化学转化、焚烧等方法进行处理，由于垃圾回收时一般不会分类，所以处理时不同固体间可能会产生剧烈的生化反应，有时还会由于处理环境相对封闭额外的产生沼气，沼气易燃且与空气混合后易爆炸可能会损坏处理设备，并且沼气的主要成分是甲烷其造成的温室效应比二氧化碳还要大得多，而且甲烷比空气轻易扩散，扩散后难集中处理。分类回收就意味着识别垃圾，具有相同属性的垃圾集中收集。气体垃圾一般不需要识别，只需要检验气体是否有毒，为了降低成本，可以不同气体集中收集，不需要识别气体的种类而分类收集。目前气体垃圾的处理主要依靠专业的防化部队，据公开资料显示中国共有13个防化团(营)，而垃圾处理的场所却很多使用防化部队处理气体垃圾的成本非常高、可行性极低，基本无法实现普及。本发明可以通过实时检测气体毒性，从而实现气体垃圾安全集中回收的普及。
综上所述，现有技术存在的问题是：
(1)现有神经网络遗漏量子计算中复数里的信息，非常容易导致层间溢出。
(2)现有的固体垃圾处理方法处理时不同固体间可能会产生剧烈的反应，产生额外的有毒气体，对人体和环境造成损害。
(3)现有的气体垃圾处理方法成本非常高，需要专业的防化部队。
解决上述技术问题的难度：固体垃圾处理对人体和环境造成伤害，气体垃圾处理成本非常高。
解决上述技术问题的意义：保护人的安全、避免污染环境、美化园区景观，大幅度降低处理成本、使普及成为可能，保留完整的事物信息，使人工神经网络避免层间溢出。
发明内容
针对现有技术存在的问题，本发明提供了一种基于次卷积超相关的垃圾识别分类处理深度学习方法。
本发明是这样实现的，一种基于次卷积超相关的垃圾识别分类处理深度学习方法包括：
第一步，深度学习卷积层的输入一组序列，根据广义二项式公式，得到序列的广义二项式以伽马函数表达的系数；
第二步，当输入的数据用次卷积做运算后，与卷积层的权重矩阵做卷积；
第三步，得到的相应输出数据，做超方差运算；
第四步，得到摄像头视野中的物体是否为矿场垃圾，并且知道该物体是哪一类垃圾；
第五步，所得到的信息传递给控制模块，控制模块触发机器手回收物体，或者就地清洗，完成整个回收与处理过程。
进一步，所述第一步具体包括：深度学习卷积层的输入为一组序列x 1，x 2，……x n，其欧几里德范数为：
Figure PCTCN2019122533-appb-000001
其中||x||≥0，x∈E，m为正整数，||x|| m为整数型范数；
当
Figure PCTCN2019122533-appb-000002
r∈[1，+∞)，扩展的分数型欧几里德范数为
Figure PCTCN2019122533-appb-000003
Figure PCTCN2019122533-appb-000004
Figure PCTCN2019122533-appb-000005
当所有的a大于0，且a≠1，x∈R，a x可以被定义为：
Figure PCTCN2019122533-appb-000006
带入扩展后的欧几里德范数，能够得到分数范数：
Figure PCTCN2019122533-appb-000007
Figure PCTCN2019122533-appb-000008
当r的范围在
Figure PCTCN2019122533-appb-000009
时，分数范数扩展到实数域。根据广义二项式公式，可以得到序列的广义二项式系数：
Figure PCTCN2019122533-appb-000010
最后，序列的F阶中心矩为：
Figure PCTCN2019122533-appb-000011
当k＝1.5时，μ 1.5为深度学习卷积层的输入做次卷积运算。
进一步，所述第二步当输入的数据用次卷积做运算后，与卷积层的权重矩阵w做卷积。
进一步，所述第三步得到的相应输出数据y 1，y 2......y n，做超方差运算，即当k＝2.5时的值。
本发明的另一目的在于提供一种基于次卷积超相关的垃圾识别分类处理深度学习系统包括：
控制系统，与深度学习模块、摄像头、无线电模块、气体识别模块、驱动模块、外设处理模块、测距模块、GPS模块相连，作用：作为单个深度学习机器人的控制中心，控制上述模块完成所有功能。控制系统由CPU、RAM、ROM、总线。
深度学习模块，与控制系统和摄像头连接，用于接受摄像头所传递过来的数据进行深度学习、并使用学习后的框架识别摄像头所拍摄的物体，判断物体属性。
摄像头，与控制系统与深度学习模块连接，用于传递机器人所行驶过的路况信息，为深度学习模块提供分析数据；
无线电模块，与控制系统连接，用于与控制中心和其它机器人通信；无线电模块主要采用工业级eSIM卡，微小、且可以远距离与控制中心联系。
气体识别模块，与控制系统连接，用于识别有毒气体并实时检测气体；气体识别模块由多种气体传感器组成。
驱动模块，与控制系统连接，用于利用电流强弱变化产生相应的磁性，为机器人提供动力；由铁质外壳与内部的磁吸组成。
外设模块，与控制系统连接，初步装备的外设是机械手臂，用于抓取固体垃圾，高级装备的外设是多模式喷淋头，用于就地清洗粉尘垃圾；
测距模块，与控制系统连接，用于发射激光束，计算机器人距离物体的距离；
GPS模块，与控制系统连接，用于获取实时位置信息。
本发明的另一目的在于提供一种与所述基于次卷积超相关的垃圾识别分类处理深度学习系统无线连接的机器人。
综上所述，本发明的优点及积极效果为：本发明垃圾定位、测距功能：当摄像头的视野中有固体垃圾时，利用激光器向机器人前方发射的激光，由光电元件接受反射3的激光束，同时计时器记录激光束从发射到接受之间的时间。利用公式L＝CT/2计算出机器人和垃圾之间的距离。其中C为激光束的速度，T为计数器记录的时间。改装的机器人使用GPS定位技术，使控制中心可以实时得到每个机器人的位置。当遇到一个机器人无法完成的任务时，控制中心根据每个机器人的位置，控制相邻机器人抵达，相互合作完成。
本发明的深度学习识别、分类处理垃圾：当摄像头视野中出现物体时，控制模块触 发深度学习模块，快速的识别物体，并对其进行分类处理。
本发明的识别有毒气体：利用气体传感器，当识别到有毒气体时，控制模块发送有毒气体位置消息给指挥中心，为后续工作提供位置信息。
本发明与已有的技术相比，优点在于：(1)在收集到一定物体时，遇到不能通行的空间时，取出收集装置里的物体，同时定位当时的位置，收缩外表。返程时带上当时放下的物体。(2)利用了新的算法改进了传统的深度学习，使识别物体的时间减少，并且在分析物体的种类方面更加准确，不会把垃圾当花朵，也不会把花朵当垃圾。(3)电流强弱的变化改变了磁力的大小，为机器人的移动提供了动力。(4)指挥中心可以控制多个机器人同时工作，减少了收集大难度。
本发明为改进的深度学习方法，比一般深度学习运行更快。本发明将深度学习识别分类功能集成到垃圾机器人中。当机器人的摄像头扫描到垃圾时，自动触发深度学习识别模式，分类到相应的类中，并利用机器手收集垃圾或就地清洗。本发明不仅可以用在环卫清洗系统，还可以用在任意人类无法达到的危险场合。
对比表
  现有垃圾处理设备 本发明所述设备
重量 5～7(吨) 1～2(吨)
大小 3～5(米) 1～3(米)
打扫场地卫生 能 能
浇花 不能 能
收集气体 不能 能
收集固体 能 能
收集液体 不能 能
AI 不具备 具备
附图说明
图1是本发明实施例提供的基于次卷积超相关的垃圾识别分类处理深度学习系统的结构示意图；
图中：1、控制系统；2、深度学习模块；3、摄像头；4、无线电模块；5、气体识别模块；6、驱动模块；7、外设处理模块；8、测距模块；9、GPS模块。
图2是本发明实施例提供的控制模块的结构示意图。
图3是本发明实施例提供的驱动模块的结构示意图。
图4是本发明实施例提供的测距模块的结构示意图。
图5是本发明实施例提供的基于次卷积超相关的垃圾识别分类处理深度学习方法流程图。
具体实施方式
为了使本发明的目的、技术方案及优点更加清楚明白，以下结合实施例，对本发明进行进一步详细说明。应当理解，此处所描述的具体实施例仅仅用以解释本发明，并不用于 限定本发明。
本发明将嵌入式技术、垃圾定位技术、信号处理及识别技术利用到矿井机器人中，实现了机器人自动识别、分类和回收处理垃圾的功能。同时又具有气体回收以及与多个机器人合作的功能。
下面结合附图对本发明的应用原理作详细的描述。
如图1-图4所示，本发明实施例提供的基于次卷积超相关的垃圾识别分类处理深度学习系统包括：控制系统1、深度学习模块2、摄像头3、无线电模块4、气体识别模块5、驱动模块6、外设处理模块7、测距模块8、GPS模块9。
控制系统1，用于连接所有相关部件
深度学习模块2，与控制系统1连接，用于分类处理不同场景
摄像头3，与深度学习模块2连接，用于传递机器人所行驶过的路况信息，为深度学习模块2提供分析数据。
无线电模块4，与控制系统1连接，用于协同周围的同伴
气体识别模块5，与控制系统1连接，用于识别有毒气体并实时检测气体。
驱动模块6，与控制系统1连接，用于利用电流强弱变化产生相应的磁性，为机器人提供动力。
外设处理模块7，与控制系统1连接，用于抓取回收或就地清洗
测距模块8，与控制系统1连接，用于发射激光束，计算机器人距离物体的距离。
GPS模块9，与控制系统1连接，用于获取实时位置信息。
本发明实施例提供的基于次卷积超相关的垃圾识别分类深度学习系统的驱动模块6利用电流强弱变化而产生相应的磁性，为机器人提供动力。当控制模块1传递有需要收集的垃圾时，无线装置发射激光束，计算机器人距离物体的距离；然后机械手从机器人的后方伸出，获取物体放入收集装置中；或者根据花卉的种类按照不同的无损角度与脉冲水压就地清洗。如果遇到较难收集的物体，指挥中心可以通过GPS模块9获取实时位置信息，控制较近的机器人共同完成。机器人采用同心圆的设计，外围为收集空间，采用可形变技术，可以自由收缩；内里为整个机器人的控制模块；气体传感器为了识别有毒气体必须实时检测气体。
如图5所示，本发明实施例提供的基于次卷积超相关的垃圾识别分类处理深度学习方法包括以下步骤：
S101：深度学习卷积层的输入一组序列，根据二项式公式，得到序列的二项式系数；
S102：当输入的数据用次卷积做运算后，与卷积层的权重矩阵做卷积；
S103：得到的相应输出数据，做超方差运算；
S104：得到摄像头视野中的物体是否为矿场垃圾，并且知道该物体是哪一类垃圾；
S105：所得到的信息传递给控制模块，控制模块触发机器手回收物体或者就地清洗，完成整个回收处理过程。
步骤S101中，根据广义二项式公式，得到序列的广义二项式以伽马函数表达的系数
本发明实施例提供的基于次卷积超相关的垃圾识别分类处理深度学习方法为了 减少深度学习识别物体的时间并且增加识别的准确性，采用了一种改进的卷积神经网络。其识别方法为：
第一步，深度学习卷积层的输入为一组序列x 1，x 2，……x n，其欧几里德范数为：
Figure PCTCN2019122533-appb-000012
其中||x||≥0，x∈E，m为正整数，||x|| m为整数型范数；
当
Figure PCTCN2019122533-appb-000013
r∈[1，+∞)，扩展的分数型欧几里德范数为
Figure PCTCN2019122533-appb-000014
Figure PCTCN2019122533-appb-000015
Figure PCTCN2019122533-appb-000016
当所有的a大于0，且a≠1，x∈R，a x可以被定义为：
Figure PCTCN2019122533-appb-000017
带入扩展后的欧几里德范数，能够得到分数范数：
Figure PCTCN2019122533-appb-000018
Figure PCTCN2019122533-appb-000019
当r的范围在
Figure PCTCN2019122533-appb-000020
时，分数范数扩展到实数域。根据广义二项式公式，可以得到序列的广义二项式系数：
Figure PCTCN2019122533-appb-000021
最后，序列的F阶中心矩为：
Figure PCTCN2019122533-appb-000022
当k＝1.5时，μ 1.5为深度学习卷积层的输入做次卷积运算。
第二步，当输入的数据用次卷积做运算后，与卷积层的权重矩阵w做卷积。
第三步，得到的相应输出数据y 1，y 2......y n，做超方差运算，即当k＝2.5时公式1的值。
第四步，可以得到摄像头视野中的物体是否为矿场垃圾，并且知道该物体是哪一类垃圾。
第五步，所得到的信息传递给控制模块，控制模块触发机器手回收或清洗物体，完成整个回收处理过程。
表1：
ARCHITEC，TURE FS PARAMS AP，％
SHALLOW，REAL 11KHZ   66.1
SHALLOW，COMPLEX 11KHZ   66.0
SHALLOW，THICKSTUETAL.(2016) 44.1KHZ   67.8
DEEP，REAL 11KHZ 10.0M 69.6
DEEP，COMPLEX 11KHZ 8.8M 72.9
表1证明在高维的深度卷积中加入复数运算，可以有效提高识别正确率。
以上所述仅为本发明的较佳实施例而已，并不用以限制本发明，凡在本发明的精神和原则之内所作的任何修改、等同替换和改进等，均应包含在本发明的保护范围之内。

Claims (7)

  1. 一种基于次卷积超相关的垃圾识别分类处理深度学习方法，其特征在于，所述基于次卷积超相关的垃圾识别分类处理深度学习方法包括：
    第一步，深度学习卷积层的输入一组序列，根据广义二项式公式，得到序列的广义二项式以伽马函数表达的系数；
    第二步，当输入的数据用次卷积做运算后，与卷积层的权重矩阵做卷积；
    第三步，得到的相应输出数据，做超方差运算；
    第四步，得到摄像头视野中的物体是否为矿场垃圾，并且获取物体属于哪一类垃圾；
    第五步，所得到的信息传递给控制模块，控制模块触发机器手回收物体，完成整个回收或就地处理过程。
  2. 如权利要求1所述的基于次卷积超相关的垃圾识别分类处理深度学习方法，其特征在于，所述第一步具体包括：深度学习卷积层的输入为一组序列x 1，x 2，……x n，欧几里德范数为：
    Figure PCTCN2019122533-appb-100001
    其中||x||≥0，x∈E，m为正整数，||x|| m为整数型范数；
    当
    Figure PCTCN2019122533-appb-100002
    r∈[1，+∞)，扩展的分数型欧几里德范数为
    Figure PCTCN2019122533-appb-100003
    Figure PCTCN2019122533-appb-100004
    Figure PCTCN2019122533-appb-100005
    当所有的a大于0，且a≠1，x∈R，a x可以被定义为：
    Figure PCTCN2019122533-appb-100006
    带入扩展后的欧几里德范数，能够得到分数范数：
    Figure PCTCN2019122533-appb-100007
    Figure PCTCN2019122533-appb-100008
    当r的范围在
    Figure PCTCN2019122533-appb-100009
    时，分数范数扩展到实数域；根据广义二项式公式，得到序列的广义二项式系数：
    Figure PCTCN2019122533-appb-100010
    最后，x不是整数，普通二项式系数的阶乘运算必须由伽马函数替代，为广义二项式，序列的F阶中心矩为：
    Figure PCTCN2019122533-appb-100011
    当k＝1.5时，μ 1，5为深度学习卷积层的输入做次卷积运算。
  3. 如权利要求1所述的基于次卷积超相关的垃圾识别分类处理深度学习方法，其特征 在于，所述第二步当输入的数据用次卷积做运算后，与卷积层的权重矩阵w做卷积。
  4. 如权利要求1所述的基于次卷积超相关的垃圾识别分类处理深度学习方法，其特征在于，所述第三步得到的相应输出数据y 1，y 2......y n，做超方差运算，当k＝2.5时的值。
  5. 一种实施权利要求1所述基于次卷积超相关的垃圾识别分类处理深度学习方法的基于次卷积超相关的垃圾识别分类处理深度学习系统，其特征在于，所述基于次卷积超相关的垃圾识别分类处理深度学习系统包括：
    控制系统，与深度学习模块、摄像头、无线电模块、气体识别模块、驱动模块、外设处理模块、测距模块、GPS模块通过CPU、RAM、ROM或总线相连，用于控制深度学习机器人并控制上述模块完成所有功能；
    深度学习模块，与控制系统和摄像头连接，用于接受摄像头所传递过来的数据进行深度学习、并使用学习后的框架识别摄像头所拍摄的物体，判断物体属性；
    摄像头，与控制系统与深度学习模块连接，用于传递机器人所行驶过的路况信息，为深度学习模块提供分析数据；
    无线电模块，与控制系统连接，用于与控制中心和其它机器人通信；
    气体识别模块，与控制系统连接，通过多种气体传感器识别有毒气体并实时检测气体；
    驱动模块，与控制系统连接，用于利用电流强弱变化产生相应的磁性，为机器人提供动力；
    外设处理模块，与控制系统连接，初步装备的外设是机械手臂，用于抓取固体垃圾；
    测距模块，与控制系统连接，用于发射激光束，计算机器人距离物体的距离；
    GPS模块，与控制系统连接，用于获取实时位置信息。
  6. 如权利要求5所述的基于次卷积超相关的垃圾识别分类处理深度学习系统，其特征在于，驱动模块包括铁质外壳及安装在铁质外壳内部的磁吸；
    外设处理模块包括多模式喷淋头，用于就地清洗粉尘垃圾。
  7. 一种与权利要求5所述基于次卷积超相关的垃圾识别分类处理深度学习系统无线连接的机器人。
PCT/CN2019/122533 2019-04-28 2019-12-03 一种基于次卷积超相关的垃圾识别分类处理深度学习方法 WO2020220674A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910349274.4 2019-04-28
CN201910349274.4A CN110059767A (zh) 2019-04-28 2019-04-28 一种基于次卷积超相关的垃圾识别分类处理深度学习方法

Publications (1)

Publication Number Publication Date
WO2020220674A1 true WO2020220674A1 (zh) 2020-11-05

Family

ID=67319475

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/122533 WO2020220674A1 (zh) 2019-04-28 2019-12-03 一种基于次卷积超相关的垃圾识别分类处理深度学习方法

Country Status (2)

Country Link
CN (1) CN110059767A (zh)
WO (1) WO2020220674A1 (zh)

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110059767A (zh) * 2019-04-28 2019-07-26 宿迁海沁节能科技有限公司 一种基于次卷积超相关的垃圾识别分类处理深度学习方法
WO2021022475A1 (zh) * 2019-08-06 2021-02-11 中国长城科技集团股份有限公司 一种垃圾处理方法、装置及终端设备
CN110694934A (zh) * 2019-09-01 2020-01-17 阿尔飞思(昆山)智能物联科技有限公司 智能干垃圾分类云端系统及其工作方法
CN113401543B (zh) * 2021-06-17 2022-07-08 广东机电职业技术学院 一种基于无人驾驶小车的垃圾分类装置和分类方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105772407A (zh) * 2016-01-26 2016-07-20 耿春茂 一种基于图像识别技术的垃圾分类机器人
CN106000904A (zh) * 2016-05-26 2016-10-12 北京新长征天高智机科技有限公司 一种生活垃圾自动分拣系统
CN108994855A (zh) * 2018-08-15 2018-12-14 深圳市烽焌信息科技有限公司 垃圾定期清理方法及机器人
CN110059767A (zh) * 2019-04-28 2019-07-26 宿迁海沁节能科技有限公司 一种基于次卷积超相关的垃圾识别分类处理深度学习方法

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105787506A (zh) * 2016-01-26 2016-07-20 耿春茂 基于图像识别与二维码识别技术的垃圾分类评价方法
CN108415419A (zh) * 2018-01-24 2018-08-17 达闼科技（北京）有限公司 垃圾拾取方法、机器人、电子设备和计算机程序产品

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105772407A (zh) * 2016-01-26 2016-07-20 耿春茂 一种基于图像识别技术的垃圾分类机器人
CN106000904A (zh) * 2016-05-26 2016-10-12 北京新长征天高智机科技有限公司 一种生活垃圾自动分拣系统
CN108994855A (zh) * 2018-08-15 2018-12-14 深圳市烽焌信息科技有限公司 垃圾定期清理方法及机器人
CN110059767A (zh) * 2019-04-28 2019-07-26 宿迁海沁节能科技有限公司 一种基于次卷积超相关的垃圾识别分类处理深度学习方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
QING ZOU , YUAN HU , JUD STEED HUANG: "Definition of Complex Hurst and Fractional Analysis for Stock Market Fluctuation", 31 May 2015 (2015-05-31), XP055750072, Retrieved from the Internet <URL:https://www.researchgate.net/publication/272680470> *

Also Published As

Publication number Publication date
CN110059767A (zh) 2019-07-26

Similar Documents

Publication Publication Date Title
WO2020220674A1 (zh) 一种基于次卷积超相关的垃圾识别分类处理深度学习方法
CN206170139U (zh) 基于无线控制的智能机器人
Bruemmer et al. A robotic swarm for spill finding and perimeter formation
Duro et al. Advances in intelligent robotics and collaborative automation
CN205809598U (zh) 一种基于物联网的水利自动化管理平台
CN109283296A (zh) 一种无人机环境监测系统及其应用
CN107422338A (zh) 一种基于北斗卫星定位的小区无人环卫云系统
CN108263777A (zh) 一种基于大数据和物联网智能垃圾分类回收方法及其装置
CN104503454A (zh) 基于多智能体理论的搜救机器人系统运动控制方法
CN105277593A (zh) 一种基于移动机器人的室内气味源定位方法
CN111070180A (zh) 基于ros的灾后救援通道检测机器人
CN113377111A (zh) 一种无人清扫车的任务调度系统及方法
CN107065892A (zh) 无线控制的灾后生命探测及环境数据采集搜救车
CN117288094B (zh) 基于激光传感器的掘进机实时定位系统
CN204331465U (zh) 一种进行复杂环境探测的机器系统
CN212471506U (zh) 一种巡检机器人
CN107440635A (zh) 一种楼宇幕墙清洁机器人清洁路径规划系统
CN211541234U (zh) 一种服务铁路客站建设及运维全流程的智能机器人
CN103744055B (zh) 一种矿区防盗采监测定位方法及其设备
CN208358811U (zh) 一种污水管线检测机器人系统
CN111984007A (zh) 自动驾驶智能扫路人机交互控制系统及设备
CN104458908A (zh) 一种基于遥控飞机的检测装置及检测方法
Permana et al. Integrated waste management system with IOT-based centralized control towards a smart eco campus-Telkom University
CN113386147B (zh) 一种基于对象识别的语音系统
Mallikarathne et al. Design a Beach Cleaning Robot Based on AI and Node-RED Interface for Debris Sorting and Monitor the Parameters

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: 19927129

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: 19927129

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 19927129

Country of ref document: EP

Kind code of ref document: A1