CN109583571A - A kind of soft ground passability prediction technique of mobile robot based on LSTM network - Google Patents
A kind of soft ground passability prediction technique of mobile robot based on LSTM network Download PDFInfo
- Publication number
- CN109583571A CN109583571A CN201811480067.4A CN201811480067A CN109583571A CN 109583571 A CN109583571 A CN 109583571A CN 201811480067 A CN201811480067 A CN 201811480067A CN 109583571 A CN109583571 A CN 109583571A
- Authority
- CN
- China
- Prior art keywords
- passability
- mobile robot
- lstm
- feature
- soft ground
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- 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
- 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
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- 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
- G06N3/08—Learning methods
Abstract
The present invention discloses a kind of soft ground passability prediction technique of the mobile robot based on LSTM network, include the following steps: 1, real-time measurement and records three kinds of passability designation dates of mobile robot, input data including traction coeficient, drive efficiency, longitudinal velocity, as model;Meanwhile the ground under the current working artificially observed can be by spending situation, label as input data;2, the soft ground passability prediction model based on LSTM building unit towards mobile robot, and a large amount of tape label data of previous step are sent into this model and are trained;3, it the model parameter in set-up procedure 2 and repeatedly trains, until obtaining the passability prediction model of stable convergence.By the above-mentioned means, the model established can be merged newly input under current working three indexs, feature extraction, and corresponding passability degree predicted values are provided, to judge whether there is a situation where retard motions for soft terrain.
Description
Technical field
The present invention relates to a kind of soft terrain passability prediction techniques of mobile robot, belong to robotically-driven kinematics
With dynamics research field.
Background technique
The soft ground passability of mobile robot (trolley) refers to that it can pass through soft surface with certain speed
The ability of (such as miriness, beach, sand ground and snowfield).Currently, wheeled robot is widely used to its unique superiority
In celestial body detecting, post-disaster search and rescue, ground transport and geological prospecting, but mellow soil or ground make robot easily occur to sink
It falls into, trackslip, reduce the traction property or dynamic property of robot, and then generate movement or turn to and be limited, can not even move
The phenomenon that passing through failure seriously hinders its practical application.Therefore, the evaluation of soft ground passability and prediction of mobile robot,
It has become increasingly by a problem concerned by people, only the design of cross-country traveling ability mobile robot and improvement do not provide one kind
Evaluation measures, while can also avoid retard motion in actual travel or reduce the probability of its generation, to ensure that robot is living
Dynamic goes on smoothly.
For the soft ground passability problem of mobile robot, most of existing method is also in passability problem evaluation
In research: as used using traction coeficient as single ground passability evaluation index, to measure robot on the ground
Driving power;For another example using drive efficiency as single evaluation index, using power loss size caused by wheel slip between
The passability judgement connect;For another example the maximum longitudinal velocity of robot motion is gone up using under certain operating condition as the characterization of dynamic property.So
And passability is not only influenced by real-time soil characteristic, also related with robot displacement state, single index can not
Passage capacity or the evaluation precision for measuring mobile robot completely are lower.
Simultaneously as the characteristic of different soils is different, passability of the mobile robot in Different Ground is not consistent, because
This also needs to prejudge the passability of subsequent time in real time, timely calibration and counter-measure is made by judging result, to keep away
Exempt from wheel slip, depression situation.However, existing passability index is still static, discrete assessment, can not accomplish in real time
Prediction and judgement, it is difficult to accurately provide in a period of time it is specific can by spend situation, can not veritably avoid retard motion,
The handling capacity of robot is improved, this technical problem must be just captured.
Summary of the invention
In order to solve the problems in background technique, the present invention provides one kind to be based on LSTM(Long Short-Term
Memory) the soft ground passability prediction technique of the mobile robot of recurrent neural network can comprehensively consider mobile robot
Dynamic property, flexibility, under the conditions of economy, accurately predict one section of moment inner machine people whether can by soft terrain,
Realized under different soil with this, can accurately estimate can by function.
In order to solve the above technical problems, the technology of the present invention solution is as follows:
A kind of soft ground passability prediction technique of mobile robot based on LSTM network, includes the following steps:
Step 1: three kinds of passability achievement datas of real-time measurement and record mobile robot, including traction coeficient, driving effect
Rate, longitudinal velocity, using them as the input data of model;Meanwhile the ground under the current working artificially observed can degree of passing through
Situation, as the label of input data;
Step 2: being based on LSTM unit, constructs the soft ground passability Network Prediction Model towards mobile robot;And by step
A large amount of tape label data obtained in one are sent into this model and are trained;
Step 3: model parameter in set-up procedure two is simultaneously repeatedly trained, until obtaining the passability prediction model of stable convergence;
The model will be according to three indexs newly inputted, and providing accordingly can be by spending predicted value.
As the further technical solution of the present invention, the measurement or calculation method of three kinds of passability indexs in the step 1
As follows: traction coeficient π is defined as the draw bar load under robot Unit Weight, i.e.,, wherein D and FZRespectively move
The draw bar load and vertical load of mobile robot, they can be surveyed by the wheel force being installed on robotically-driven wheel
?;Drive efficiency e is defined as the ratio between the output power of motor and input power on driving wheel, i.e.,, wherein driving is turned round
Square T is directly measured by wheel force, angular speed of wheelIt can be obtained by the rotary encoder built in wheel force defeated
Differential calculation acquisition is carried out out;The forward speed V of the horizontal aspect of mobile robot is longitudinal velocity, can be by being installed on machine
Inertial Measurement Unit on people's center chassis position provides in real time.
As the further technical solution of the present invention, the acquisition of three kinds of passability achievement datas in the step 1 have with
Lower technical characteristic: identical time window should be taken, and in a time windowTInterior data are as an actual samples sample
This, sample frequency is set as 1Hz, and enough sample datas are acquired under Current terrestrial operating condition, and sample number is greater than 2000, with
Guarantee the accuracy of model;The soft ground of the robot artificially observed can use 0 or 1 form coding by degree, i.e. robot can
Be denoted as 1 by the ground, cannot by (as occur depression, skid) if be denoted as 0.
As the further technical solution of the present invention, the passability Network Prediction Model in the step 2, which uses, is based on LSTM
The deep learning algorithm frame of module, the algorithm framework specific steps are as follows:
Step 2.1, passability achievement data obtained in step 1 is pre-processed, by each moment t in each sample
One LSTM unit of data access, each LSTM unit be contain forget door, input gate, out gate standard recurrent neural net
Network module;
Step 2.2, a large amount of LSTM units in step 2.1 are designed as double-layer structure form LSTM network, first layer LSTM net
Network is for primary feature extraction, and the validity feature that the second layer carries out deep layer extracts, and three passability achievement datas will be right respectively
The mutually isostructural LSTM network of Ying Yusan group;
Step 2.3, the characteristic that step 2.2 obtains each index is accessed into a cascading layers, this layer will each refer to target feature
Value carries out rearranging integration, obtains fused feature vector;
Step 2.4, feature vector step 2.3 obtained accesses a full articulamentum, carries out further arameter optimization and place
Reason;
Step 2.5, the output data of step 2.4 is accessed into a Softmax and returns classifier, which is used for characteristic value
It is transformed to can be used for classifying 0 or 1 output, realizes the function of passability prediction.
As the further technical solution of the present invention, the LSTM unit used in the sub-step 2.1 of the step 2 can root
Judge whether the information of input is useful, and core is the state c with ' door ' come control unit according to rule.In current t moment, lose
Forget a ftIt is responsible for the c of control last momentt-1C of how many information preservation to current timet;Input gate itSelect current time
Immediate status ct *How many information input is to active cell;Out gate otControl current state ctHow many information is as the moment
Hidden layer export ht;Instant alternative state ct *By the input x of current t moment networktH is exported with the hidden layer of last momentt-1Jointly
It determines, their calculation formula is respectively
Wherein, W and b is respectively weight matrix and bias term, subscriptf、i、o、cIt respectively represents and forgets door, input gate, output
Door, alternative state;Activation for activation primitive, ' door ' state uses sigmoid function, immediate status ct *Activation adopt
With tanh function.Finally, the hidden layer of current time LSTM unit exports htBy out gate otWith active cell state ct
It codetermines;Current state ctOutput can be by forgetting door ft, last moment state ct-1, input gate it, immediate status ct *Jointly
It determines;And LSTM unit is exported in the immediate status of subsequent timeH can be exported by hidden layert, output layer weight matrix W, with
And bias term b is codetermined, their calculation formula is respectively
In formula,Symbol expression is multiplied by element.
Characteristic value, which is obtained, after cascading layers fusion in the sub-step 2.3 of the step 2 uses one-dimensional characteristic vector form,
It indicates are as follows:
In formula,Respectively indicate corresponding to the traction coeficient characteristic value of i-th of sample, drive efficiency feature,
Longitudinal velocity feature, they are scalar value.
As the further technical solution of the present invention, the full articulamentum in the sub-step 2.4 of the step 2 is exported each
Moment, t was independent from each other, and in each moment t, the output of full articulamentum is equally represented byForm, wherein
XtInput from preceding layer.
As the further technical solution of the present invention, Softmax returns classifier calculated in the sub-step 2.5 of the step 2
The probability Estimation formula of j-th of classification is
In formula,Indicate the matrix being made of two classification parameters, the value of classification j is 1 or 2.At this point,Table
Show the input feature vector value of i-th of sample from a upper layer network,Whether pass through for robot under i-th of sample
Label value, they can be denoted as sample pair jointly。
As the further technical solution of the present invention, in the LSTM network training of the step 2 and step 3, the present invention is adopted
It is calculated with loss function, loss function is to return the cross entropy cost function that classifier matches with Softmax, is calculated
Formula are as follows:
In formula, N is sample number,Indicate that i-th of sample belongs to the true value of jth class,It is obtained for Softmax function category device
Predicted value.
As the further technical solution of the present invention, the model parameter adjustment in the step 3 have the feature that firstly,
Model solution method selection is momentum stochastic gradient descent method;Secondly, it is 0.1 that initial learning rate, which is arranged, initial maximum changes
It is 50 for the period;Again, ten folding cross validation modes are selected to carry out adjustment.
Mobile robot of the present invention, which refers to, is widely used on field work or mellow soil (miriness, beach, sand
Ground and snowfield) traveling wheeled mobile robot, such as common motor-driven carrier, celestial body detecting vehicle, desert test platform, Yi Jiwu
People's off-road vehicle etc. can be applied to front-wheel drive, four-wheel drive or the caterpillar type robot based on wheeled.
Compared with the existing technology, the beneficial effects of the present invention are:
The present invention provides a kind of soft ground passability prediction techniques of mobile robot based on LSTM network, can examine in synthesis
Under the conditions of considering the dynamic property of robot, flexibility, economy, can estimating robot in advance pass through the function of soft terrain
Can, that is, avoid soft terrain from trackslipping, depression the problems such as.Traditional mobile mechanism's passability problem is mostly using single
Evaluation index measures its handling capacity, and does not have and estimate, arbitration functions, and the present invention is by designing reasonable LSTM recurrence mind
Through network, three kinds of traction coeficient, drive efficiency, real-time speed indexs can be preferably merged, while processing continuous time in real time
Sequence data, and then providing robot can be by effective judgement of degree.
Detailed description of the invention
Fig. 1 is algorithm frame and principle of the invention.
Fig. 2 is the LSTM network model designed in step 2 of the present invention.
Fig. 3 is the structure chart of LSTM network unit of the present invention.
Fig. 4 is data acquisition schematic diagram of the mobile robot of the present invention in Different Ground.
Fig. 5 is passability model prediction result example of the mobile robot of the present invention in three kinds of soft terrains.
Specific embodiment
With reference to the accompanying drawings and detailed description, the present invention is further explained, but protection scope of the present invention is not limited to
In the case study on implementation.
As shown in Figure 1, a kind of soft ground passability prediction technique of mobile robot based on LSTM network, including walk as follows
It is rapid:
Step 1: three kinds of passability designation dates of real-time measurement and record mobile robot, including traction coeficient, driving effect
Rate, longitudinal velocity, using them as the input data of model;Meanwhile the ground under the current working artificially observed can degree of passing through
Situation, as the label of input data.
Step 2: being based on LSTM recurrent neural network unit, constructs the soft ground passability network towards mobile robot
Prediction model;And a large amount of tape label data obtained in step 1 are sent into this model and are trained.
Step 3: model parameter in set-up procedure two is simultaneously repeatedly trained, until obtaining the passability prediction of stable convergence
Model;The model will be according to three indexs newly inputted, and providing accordingly can be by spending predicted value.
The measurement of three kinds of passability indexs in the step 1 or calculation method are as follows: traction coeficient π is defined as machine
Draw bar load under people's Unit Weight, i.e.,, wherein D and FZThe respectively draw bar load of mobile robot and vertical
Load, they can be measured by the wheel force being installed on robotically-driven wheel;Drive efficiency e is defined as on driving wheel
The ratio between the output power of motor and input power, i.e.,, wherein driving torque T is directly measured by wheel force, vehicle
Take turns angular speedIt can must be exported by the rotary encoder built in wheel force and carry out differential calculation acquisition;Mobile robot
The forward speed of horizontal aspect be longitudinal velocity, can be by the Inertial Measurement Unit that is installed on robot center chassis position
It provides in real time.
In a preferable example of the invention, specific sensor data transmission and reading be can be used such as under type reality
Existing: the micro-control unit of wheel force can be defeated by the angle of the draw bar load of measurement, vertical load and encoder
Data are packaged together out, are sent to host computer by bluetooth approach;The output data of Inertial Measurement Unit then passes through the USB of itself
Serial ports and host computer Direct Communication.The LabVIEW software readable for being installed on host computer takes the data at each moment, and is adding from the background
Add waveform icon module, so that the front plate can show that the waveform diagram of numerical value.
The data acquisition of three passability indexs in the step 1 should take identical time window, and at one
Between windowTFor interior data as an actual samples sample, sample frequency is set as 1Hz, and the sample under Current terrestrial operating condition
Number is at least 2000, to guarantee the accuracy of model;The soft ground of the robot artificially observed can use 0 or 1 form by degree
Coding, i.e., robot can be denoted as 1 by the ground, cannot by (as occur depression, skidding) if be denoted as 0.
In a preferable example of the invention, in order to reach preferably effect, the data of the method for the present invention description
Format can be used arranged below: the time window T of three passability indexs is set as 1min, the input of i-th of sample at this time
Data can be expressed as
,,
The total sample number of acquisition is 4000, wherein 80% data are used for the training of network model, the data of residue 20% are tested for testing
Card.
Passability prediction model in the step 2 uses the deep learning algorithm frame based on LSTM module, the algorithm
Framework is designed according to following form, as shown in Figure 2:
Step 2.1, passability achievement data obtained in step 1 is pre-processed, by each moment t in each sample
One LSTM unit of data access, each LSTM unit be contain forget door, input gate, out gate standard recurrent neural net
Network module;
Step 2.2, a large amount of LSTM units in step 2.1 are designed as double-layer structure form LSTM network, first layer LSTM net
Network is for primary feature extraction, and the validity feature that the second layer carries out deep layer extracts, and three passability achievement datas will be right respectively
The mutually isostructural LSTM network of Ying Yusan group;
Step 2.3, the characteristic that step 2.2 obtains each index is accessed into a cascading layers, this layer will each refer to target feature
Value carries out rearranging integration, obtains fused feature vector;
Step 2.4, feature vector step 2.3 obtained accesses a full articulamentum, carries out further arameter optimization and place
Reason;
Step 2.5, the output data of step 2.4 is accessed into a Softmax and returns classifier, which is used for characteristic value
It is transformed to can be used for classifying 0 or 1 output, realizes the function of passability prediction.
As shown in figure 3, each LSTM unit can judge whether the information of input is useful, and core is to use according to rule
' door ' carrys out the state of control unit.In current t moment, door f is forgottentIt is responsible for the c of control last momentt-1How many information preservation
To the c at current timet;Input gate itSelect the immediate status c at current timet *How many information input is to active cell;Output
Door otControl current state ctHow many information exports h as the hidden layer at the momentt;Instant alternative state ct *By current t moment net
The input x of networktH is exported with the hidden layer of last momentt-1It codetermines, their calculation formula is respectively
Wherein, W and b is respectively weight matrix and bias term, subscriptf、i、o、cIt respectively represents and forgets door, input gate, output
Door, alternative state;Activation for activation primitive, ' door ' state uses sigmoid function, immediate status ct *Activation adopt
With tanh function.Finally, the hidden layer of current time LSTM unit exports htBy out gate otWith active cell state ctAltogether
With decision;Current state ctOutput can be by forgetting door ft, last moment state ct-1, input gate it, immediate status ct *It is common to determine
It is fixed;And LSTM unit is exported in the immediate status of subsequent timeH can be exported by hidden layert, output layer weight matrix W and
Bias term b is codetermined, their calculation formula is respectively
In formula,Symbol expression is multiplied by element.
Characteristic value, which is obtained, after cascading layers fusion in the sub-step 2.3 of the step 2 uses one-dimensional characteristic vector form,
It indicates are as follows:
In formula,Respectively indicate corresponding to the traction coeficient characteristic value of i-th of sample, drive efficiency feature,
Longitudinal velocity feature, they are scalar value.
Full articulamentum output in the sub-step 2.4 of the step 2 is independent from each other in each moment t, when each
T is carved, the output of full articulamentum is equally represented byForm, wherein XtInput from preceding layer.
The probability Estimation formula of Softmax function calculating the classification is in the sub-step 2.5 of the step 2
In formula,Indicate the matrix being made of two classification parameters, the value of classification j is 1 or 2.At this point,Table
Show the input feature vector value of i-th of sample from a upper layer network,Whether pass through for robot under i-th of sample
Label value, they can be denoted as sample pair jointly。
In the LSTM network training of the step 2 and step 3, loss function that the present invention uses is returns with Softmax
The cross entropy cost function for returning classifier to match, its calculation formula is:
In formula, N is sample number,Indicate that i-th of sample belongs to the true value of jth class,It is obtained for Softmax function category device
Predicted value.
Model parameter adjustment in the step 3 has the feature that firstly, model solution method selection is momentum
Stochastic gradient descent method;Secondly, it is 0.1 that initial learning rate, which is arranged, initial maximum iteration cycle is 50;Again, ten foldings are selected
Cross validation mode carries out adjustment.
In a preferable example of the invention, network algorithm model framework of the invention can be based on TensorFlow environment
Realize the building of LSTM network, hyper parameter adjust setting include: input time step-length 60, input feature vector dimension be 3, momentum it is random
Gradient uses Adam optimizer, LSTM unit input layer number 100, LSTM unit 32, Softmax node layer number for 3, uses
Portable hard platform have the central processing unit (CPU) of one piece of Intel (R) Core (TM) i7-7700, dominant frequency is
3.60 GHz, caching 16.0 GB of RAM, can satisfy training requirement.
A preferable actual test process of the invention is as shown in figure 4, since different types of ground has biggish spy
Difference is levied, the present invention has selected the validity of three kinds of different plot progress test methods, comprising dry sand ground, meadow, muddy ground, and
Mobile robot driving path is preferable along near linear;To improve actual prediction effect, the passability of mobile robot
Data acquisition can carry out data in same plot, 3 times of actual test areas of collecting training data area area under same plot
Domain area.It is illustrated in figure 5 the passability prediction result in the actual test area in three different plot;In terms of result, mobile machine
On these ground when driving, the real-time passability predictablity rate that model provides is higher by people.
The above is one embodiment of the present of invention, is not intended to restrict the invention.It is all principle of the invention it
Equivalent replacement that is interior, being done, should all be included in the protection scope of the present invention.The content that the present invention does not elaborate belongs to
Prior art well known to this professional domain technical staff.
Claims (10)
1. a kind of soft ground passability prediction technique of mobile robot based on LSTM network, which is characterized in that including following step
It is rapid:
Step 1: three kinds of passability designation dates of real-time measurement and record mobile robot, including traction coeficient, driving effect
Rate, longitudinal velocity, using them as the input data of model;Meanwhile the ground under the current working artificially observed can degree of passing through
Situation, as the label of input data;
Step 2: being based on LSTM unit, constructs the soft ground passability Network Prediction Model towards mobile robot, and by step
A large amount of tape label data obtained in one are sent into this model and are trained;
Step 3: model parameter in set-up procedure two is simultaneously repeatedly trained, until obtaining the passability prediction model of stable convergence;
The model will be according to three indexs newly inputted, and providing accordingly can be by spending predicted value.
2. the soft ground passability prediction technique of the mobile robot according to claim 1 based on LSTM network, feature
It is, the measurement of three in the step a kind passability index or calculation method are as follows: traction coeficient π is defined as robot list
Draw bar load under the weight of position, i.e.,, wherein D and FZThe respectively draw bar load of mobile robot and vertical load
Lotus;Drive efficiency e is defined as the ratio between the output power of motor and input power on driving wheel, i.e.,;Mobile robot
The forward speed V of horizontal aspect is longitudinal velocity.
3. the soft ground passability prediction technique of the mobile robot according to claim 2 based on LSTM network, feature
It is, the D and FZIt is measured by the wheel force being installed on robotically-driven wheel;Driving torque T is sensed by vehicle wheel forces
Device directly measures, angular speed of wheelProgress differential calculation must be exported by the rotary encoder built in wheel force to obtain
, longitudinal velocityvIt is provided in real time by the Inertial Measurement Unit being installed on robot center chassis position.
4. the soft ground passability prediction technique of the mobile robot according to claim 2 or 3 based on LSTM network, special
Sign is that the data acquisition of three in the step 1 passability index has the feature that data acquisition using identical
Time window, and in a time windowTInterior data definition is a sample, sample frequency 1Hz;That artificially observes leads to
Excessive label uses 0 or 1 form coding, i.e. robot can be denoted as 1 by the ground, cannot be by being then denoted as 0.
5. the soft ground passability prediction technique of the mobile robot according to claim 4 based on LSTM network, feature
It is, the passability Network Prediction Model in the step 2 uses the deep learning algorithm frame based on LSTM module, specifically
Step are as follows:
Step 2.1, passability achievement data obtained in step 1 is pre-processed, by each moment t in each sample
One LSTM unit of data access, each LSTM unit be contain forget door, input gate, out gate standard recurrent neural net
Network module;
Step 2.2, a large amount of LSTM units in step 2.1 are designed as double-layer structure form LSTM network, first layer LSTM net
Network is for primary feature extraction, and the validity feature that the second layer carries out deep layer extracts, and three passability achievement datas will be right respectively
The mutually isostructural LSTM network of Ying Yusan group;
Step 2.3, the characteristic that step 2.2 obtains each index is accessed into a cascading layers, this layer will each refer to target feature
Value carries out rearranging integration, obtains fused feature vector;
Step 2.4, feature vector step 2.3 obtained accesses a full articulamentum, carries out further arameter optimization and place
Reason;
Step 2.5, the output data of step 2.4 is accessed into a Softmax and returns classifier, which is used for characteristic value
It is transformed to can be used for classifying 0 or 1 output, realizes the function of passability prediction.
6. the soft ground passability prediction technique of the mobile robot according to claim 5 based on LSTM network, feature
It is, the characteristic value of the cascading layers fusion in the sub-step 2.3 uses one-dimensional characteristic vector form, indicates are as follows:;In formula,Respectively indicate the traction coeficient spy for corresponding to i-th of sample
Value indicative, drive efficiency feature, longitudinal velocity feature, they are scalar value.
7. the soft ground passability prediction technique of the mobile robot according to claim 5 based on LSTM network, feature
It is, the full articulamentum output in the sub-step 2.4 is independent from each other in each moment t, complete to connect in each moment t
The output of layer is equally represented byForm, wherein XtInput from preceding layer.
8. the soft ground passability prediction technique of the mobile robot according to claim 5 based on LSTM network, feature
It is, Softmax returns the probability Estimation of j-th of classification of classifier calculated in the sub-step 2.5, indicates are as follows:
;In formula,Expression is made of two classification parameters
Matrix, the value of classification j is 1 or 2,For the input feature vector of i-th of sample from a upper layer network,
The label value whether passed through for the robot of i-th of sample.
9. the soft ground passability prediction technique of the mobile robot according to claim 1 or 2 based on LSTM network, special
Sign is, in the training of the step 2 and step 3, is calculated using loss function, and loss function is to return with Softmax
The cross entropy cost function for returning classifier to match, is expressed as follows:;In formula, N is sample number,Indicate that i-th of sample belongs to the true value of jth class,The predicted value obtained for Softmax function category device.
10. the soft ground passability prediction technique of the mobile robot according to claim 1 or 2 based on LSTM network,
It is characterized in that, the model parameter adjustment in the step 3 has following characteristics: firstly, model solution method selection is momentum
Stochastic gradient descent method;Secondly, it is 0.1 that initial learning rate, which is arranged, initial maximum iteration cycle is 50;Again, ten foldings are selected
Cross validation mode carries out adjustment.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811480067.4A CN109583571B (en) | 2018-12-05 | 2018-12-05 | Mobile robot soft ground trafficability prediction method based on LSTM network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811480067.4A CN109583571B (en) | 2018-12-05 | 2018-12-05 | Mobile robot soft ground trafficability prediction method based on LSTM network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109583571A true CN109583571A (en) | 2019-04-05 |
CN109583571B CN109583571B (en) | 2023-04-28 |
Family
ID=65926399
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811480067.4A Active CN109583571B (en) | 2018-12-05 | 2018-12-05 | Mobile robot soft ground trafficability prediction method based on LSTM network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109583571B (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110362081A (en) * | 2019-07-16 | 2019-10-22 | 武昌理工学院 | A kind of method for planning path for mobile robot |
CN110390386A (en) * | 2019-06-28 | 2019-10-29 | 南京信息工程大学 | Sensitive shot and long term accumulating method based on input variation differential |
CN110472726A (en) * | 2019-07-25 | 2019-11-19 | 南京信息工程大学 | Sensitive shot and long term accumulating method based on output variation differential |
CN110490299A (en) * | 2019-07-25 | 2019-11-22 | 南京信息工程大学 | Sensitive shot and long term accumulating method based on state change differential |
CN111210091A (en) * | 2020-02-25 | 2020-05-29 | 上海积成能源科技有限公司 | System model for predicting short-term power load based on long and short-term memory model of recurrent neural network |
CN111537888A (en) * | 2020-05-09 | 2020-08-14 | 国网福建省电力有限公司莆田供电公司 | Data-driven echelon battery SOC prediction method |
CN111651935A (en) * | 2020-05-25 | 2020-09-11 | 成都千嘉科技有限公司 | Multi-dimensional expansion prediction method and device for non-stationary time series data |
CN111679970A (en) * | 2020-05-14 | 2020-09-18 | 南京理工大学 | Robot software system running environment state prediction method |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150206427A1 (en) * | 2014-01-17 | 2015-07-23 | International Business Machines Corporation | Prediction of local and network-wide impact of non-recurrent events in transportation networks |
CN106407649A (en) * | 2016-08-26 | 2017-02-15 | 中国矿业大学(北京) | Onset time automatic picking method of microseismic signal on the basis of time-recursive neural network |
CN108647839A (en) * | 2018-06-28 | 2018-10-12 | 中国人民解放军陆军军事交通学院镇江校区 | Voltage-stablizer water level prediction method based on cost-sensitive LSTM Recognition with Recurrent Neural Network |
CN108664687A (en) * | 2018-03-22 | 2018-10-16 | 浙江工业大学 | A kind of industrial control system space-time data prediction technique based on deep learning |
-
2018
- 2018-12-05 CN CN201811480067.4A patent/CN109583571B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150206427A1 (en) * | 2014-01-17 | 2015-07-23 | International Business Machines Corporation | Prediction of local and network-wide impact of non-recurrent events in transportation networks |
CN106407649A (en) * | 2016-08-26 | 2017-02-15 | 中国矿业大学(北京) | Onset time automatic picking method of microseismic signal on the basis of time-recursive neural network |
CN108664687A (en) * | 2018-03-22 | 2018-10-16 | 浙江工业大学 | A kind of industrial control system space-time data prediction technique based on deep learning |
CN108647839A (en) * | 2018-06-28 | 2018-10-12 | 中国人民解放军陆军军事交通学院镇江校区 | Voltage-stablizer water level prediction method based on cost-sensitive LSTM Recognition with Recurrent Neural Network |
Non-Patent Citations (1)
Title |
---|
魏凤: "车辆地面通过性测试系统设计", 《中国优秀硕士学位论文全文数据库工程科技Ⅱ辑》 * |
Cited By (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110390386A (en) * | 2019-06-28 | 2019-10-29 | 南京信息工程大学 | Sensitive shot and long term accumulating method based on input variation differential |
CN110362081A (en) * | 2019-07-16 | 2019-10-22 | 武昌理工学院 | A kind of method for planning path for mobile robot |
CN110362081B (en) * | 2019-07-16 | 2020-04-24 | 武昌理工学院 | Mobile robot path planning method |
CN110472726A (en) * | 2019-07-25 | 2019-11-19 | 南京信息工程大学 | Sensitive shot and long term accumulating method based on output variation differential |
CN110490299A (en) * | 2019-07-25 | 2019-11-22 | 南京信息工程大学 | Sensitive shot and long term accumulating method based on state change differential |
CN110490299B (en) * | 2019-07-25 | 2022-07-29 | 南京信息工程大学 | Sensitive long-short term memory method based on state change differential |
CN110472726B (en) * | 2019-07-25 | 2022-08-02 | 南京信息工程大学 | Sensitive long-short term memory method based on output change differential |
CN111210091A (en) * | 2020-02-25 | 2020-05-29 | 上海积成能源科技有限公司 | System model for predicting short-term power load based on long and short-term memory model of recurrent neural network |
CN111537888A (en) * | 2020-05-09 | 2020-08-14 | 国网福建省电力有限公司莆田供电公司 | Data-driven echelon battery SOC prediction method |
CN111679970A (en) * | 2020-05-14 | 2020-09-18 | 南京理工大学 | Robot software system running environment state prediction method |
CN111679970B (en) * | 2020-05-14 | 2022-09-27 | 南京理工大学 | Method for predicting running environment state of robot software system |
CN111651935A (en) * | 2020-05-25 | 2020-09-11 | 成都千嘉科技有限公司 | Multi-dimensional expansion prediction method and device for non-stationary time series data |
Also Published As
Publication number | Publication date |
---|---|
CN109583571B (en) | 2023-04-28 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109583571A (en) | A kind of soft ground passability prediction technique of mobile robot based on LSTM network | |
Zhu et al. | Forming a new small sample deep learning model to predict total organic carbon content by combining unsupervised learning with semisupervised learning | |
CN103364006B (en) | For determining the system and method for vehicle route | |
Wang et al. | On a new method of estimating shear wave velocity from conventional well logs | |
Yang et al. | Convolutional neural network–based friction model using pavement texture data | |
Ward et al. | Speed-independent vibration-based terrain classification for passenger vehicles | |
CN108664687A (en) | A kind of industrial control system space-time data prediction technique based on deep learning | |
CN110184885B (en) | Method for testing pavement evenness based on smart phone | |
Zhang et al. | Application of a machine learning method to evaluate road roughness from connected vehicles | |
Wang et al. | A hybrid framework for reservoir characterization using fuzzy ranking and an artificial neural network | |
CN106324674B (en) | A kind of shale gas TOC pre-stack seismic inversion Forecasting Methodologies | |
CN105760673A (en) | Fluvial facies reservoir earthquake sensitive parameter template analysis method | |
Deng et al. | Upscaling retardation factor in hierarchical porous media with multimodal reactive mineral facies | |
Shoop et al. | Overview of cold regions mobility modeling at CRREL | |
Shi et al. | A new method for rock brittleness evaluation in tight oil formation from conventional logs and petrophysical data | |
CN110509916A (en) | A kind of body gesture antihunt means and system based on deep neural network | |
Tolle et al. | Lacunarity definition for ramified data sets based on optimal cover | |
CN103454677A (en) | Seismic data retrieval method based on combination of particle swarm and linear adder | |
Liu et al. | Road roughness acquisition and classification using improved restricted Boltzmann machine deep learning algorithm | |
CN108256238A (en) | A kind of optic fiber grating wavelength demodulation method and device based on deep learning | |
Zhao et al. | Lithofacies identification of shale reservoirs using a tree augmented Bayesian network: A case study of the lower Silurian Longmaxi formation in the changning block, South Sichuan basin, China | |
Habibnejad Korayem et al. | Road angle estimation for a vehicle-trailer with machine learning and system model-based approaches | |
CN108228994B (en) | Stress calculation method for vehicle and working device under random excitation of off-road | |
Hsieh et al. | Parameterization of cloud droplet size distributions: Comparison with parcel models and observations | |
Guo et al. | A mobile sensing system for urban monitoring with adaptive resolution |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |