CN109583571B - Mobile robot soft ground trafficability prediction method based on LSTM network - Google Patents
Mobile robot soft ground trafficability prediction method based on LSTM network Download PDFInfo
- Publication number
- CN109583571B CN109583571B CN201811480067.4A CN201811480067A CN109583571B CN 109583571 B CN109583571 B CN 109583571B CN 201811480067 A CN201811480067 A CN 201811480067A CN 109583571 B CN109583571 B CN 109583571B
- Authority
- CN
- China
- Prior art keywords
- trafficability
- mobile robot
- data
- lstm
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 45
- 238000012549 training Methods 0.000 claims abstract description 14
- 230000006870 function Effects 0.000 claims description 21
- 238000004364 calculation method Methods 0.000 claims description 11
- 239000011159 matrix material Substances 0.000 claims description 9
- 239000013598 vector Substances 0.000 claims description 9
- 238000000605 extraction Methods 0.000 claims description 6
- 238000013528 artificial neural network Methods 0.000 claims description 5
- 230000000306 recurrent effect Effects 0.000 claims description 5
- 238000005070 sampling Methods 0.000 claims description 5
- 238000005259 measurement Methods 0.000 claims description 4
- 238000012545 processing Methods 0.000 claims description 4
- 238000012795 verification Methods 0.000 claims description 4
- 238000013480 data collection Methods 0.000 claims description 3
- 238000013135 deep learning Methods 0.000 claims description 3
- 230000004927 fusion Effects 0.000 claims description 3
- 238000011478 gradient descent method Methods 0.000 claims description 3
- 238000007781 pre-processing Methods 0.000 claims description 3
- 230000033001 locomotion Effects 0.000 abstract description 6
- 238000011156 evaluation Methods 0.000 description 8
- 239000002689 soil Substances 0.000 description 5
- 230000004913 activation Effects 0.000 description 4
- 238000012360 testing method Methods 0.000 description 4
- 238000001514 detection method Methods 0.000 description 3
- 239000004576 sand Substances 0.000 description 3
- 238000010586 diagram Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 239000004973 liquid crystal related substance Substances 0.000 description 2
- 238000000691 measurement method Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
Images
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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a mobile robot soft ground trafficability prediction method based on an LSTM network, which comprises the following steps: 1. three types of trafficability indicating data of the mobile robot are measured and recorded in real time, wherein the three types of trafficability indicating data comprise traction coefficients, driving efficiency and longitudinal speed, and the traction coefficients, the driving efficiency and the longitudinal speed are used as input data of a model; meanwhile, the ground passing condition under the current working condition of artificial observation is used as a tag for inputting data; 2. constructing a soft ground trafficability prediction model for the mobile robot based on the LSTM unit, and sending a large amount of labeled data of the previous step into the model for training; 3. and (3) adjusting the model parameters in the step (2) and training for multiple times until a stable convergent pass prediction model is obtained. By the mode, the built model can fuse and extract the three newly input indexes under the current working condition and give out corresponding trafficability degree predicted values, so that whether the movement of the soft ground is blocked or not is judged.
Description
Technical Field
The invention relates to a soft ground trafficability prediction method of a mobile robot, and belongs to the field of robot driving kinematics and dynamics research.
Background
The soft ground trafficability of a mobile robot (trolley) refers to its ability to smoothly pass soft road surfaces (such as mud, mud flat, sand, snow, etc.) at a certain speed. At present, the wheel type robot has been widely applied to star detection, post-disaster search and rescue, ground transportation and geological exploration by virtue of the unique superiority, but soft soil or ground enables the robot to easily sink, slip and the like, so that the traction or dynamic property of the robot is reduced, and further, the phenomenon of passing failure such as limited movement or steering and even incapability of moving is generated, so that the practical application of the wheel type robot is seriously hindered. Therefore, soft ground trafficability evaluation and prediction of the mobile robot become a problem which is more and more focused, and not only provides an evaluation means for designing and improving the high-trafficability mobile robot, but also can avoid the blockage of the movement or reduce the occurrence probability of the movement in the actual running process, thereby ensuring the smooth running of the robot.
Aiming at the soft ground trafficability problem of the mobile robot, most of the existing methods are still in trafficability problem evaluation research: if the traction coefficient is used as a single ground trafficability evaluation index, the traction capacity of the robot on the ground is measured; for example, the driving efficiency is used as a single evaluation index, and the power loss caused by wheel slip is used as indirect trafficability judgment; and the maximum longitudinal speed of the movement of the upper robot under a certain working condition is used as the dynamic characteristic. However, the trafficability is not only affected by the real-time soil characteristics, but also related to the motion state of the robot itself, and a single index cannot fully measure the trafficability of the mobile robot or the evaluation accuracy is low.
Meanwhile, as the characteristics of different soils are different, the trafficability of the mobile robot on different grounds is not consistent, so that the trafficability of the mobile robot at the next moment is also required to be pre-judged in real time, and timely calibration and countermeasure measures are made according to the judging result so as to avoid the wheel slip and subsidence condition. However, the conventional trafficability index is still static and discrete to evaluate, and cannot be predicted and judged in real time, so that a specific trafficability condition in a period of time cannot be accurately given, and the technical problem must be solved in order to improve the trafficability of the robot.
Disclosure of Invention
In order to solve the problems in the background art, the invention provides a soft ground trafficability prediction method of a mobile robot based on an LSTM (Long Short-Term Memory) recurrent neural network, which can accurately predict whether the robot can pass through soft ground within a period of time under the condition of comprehensively considering the dynamic property, flexibility and economy of the mobile robot, so that the function of accurately predicting whether the robot can pass through under different soil conditions is realized.
In order to solve the technical problems, the technical scheme of the invention is as follows:
a mobile robot soft ground trafficability prediction method based on an LSTM network comprises the following steps:
step one: three types of trafficability index data of the mobile robot, including traction coefficient, driving efficiency and longitudinal speed, are measured and recorded in real time and used as input data of a model; meanwhile, the ground passing condition under the current working condition of artificial observation is used as a label of input data;
step two: constructing a soft ground trafficability network prediction model facing the mobile robot based on the LSTM unit; sending a large amount of labeled data obtained in the first step into the model for training;
step three: adjusting the model parameters in the second step and training for multiple times until a stable convergent passing prediction model is obtained; the model can give corresponding pass degree predicted values according to the three indexes which are newly input.
As a further technical scheme of the present invention, the measurement or calculation method of the three trafficability indexes in the first step is as follows: the traction coefficient pi is defined as the traction force of the hook per unit weight of the robot, i.eWherein D and F Z The traction force and the vertical load of the hook of the mobile robot are respectively measured by a wheel force sensor arranged on a driving wheel of the robot; the driving efficiency e is defined as the ratio of the output power to the input power of the motor on the driving wheel, i.e. +.>Wherein the driving torque T is directly measured by the wheel force sensor, the wheel angular speed +.>The output of the wheel force sensor is obtained through differential calculation by a rotary encoder arranged in the wheel force sensor; the horizontal forward speed V of the mobile robot is the longitudinal speed, and can be given in real time by an inertial measurement unit installed at the center of the chassis of the robot.
As a further technical scheme of the invention, the collection of the three trafficability index data in the first step has the following technical characteristics: the same time window should be taken and within one time windowTThe data in the model is taken as an actual sampling sample, the sampling frequency is set to be 1Hz, and enough sample data are collected under the current ground working condition, and the number of the samples is more than 2000 so as to ensure the accuracy of the model; the trafficability of the soft ground of the robot observed by people adopts a form of 0 or 1 to code, namely the robot can be marked as 1 through the ground, and the robot can not pass through (such as sinking and slipping) and is marked as 0.
As a further technical scheme of the invention, the pass-through network prediction model in the second step adopts a deep learning algorithm framework based on an LSTM module, and the algorithm framework comprises the following specific steps:
step 2.1, preprocessing the trafficability index data obtained in the step one, and accessing the data of each time t in each sample into an LSTM unit, wherein each LSTM unit is a standard recurrent neural network module comprising a forgetting gate, an input gate and an output gate;
step 2.2, designing a large number of LSTM units in the step 2.1 into a double-layer structure type LSTM network, wherein the first layer LSTM network is used for primary feature extraction, the second layer is used for deep effective feature extraction, and three pieces of trafficability index data respectively correspond to three groups of LSTM networks with the same structure;
step 2.3, the characteristic data of each index obtained in the step 2.2 are accessed into a cascade layer, and the characteristic values of the respective indexes are rearranged and integrated by the cascade layer to obtain a fused characteristic vector;
step 2.4, the feature vector obtained in the step 2.3 is accessed into a full connection layer, and further parameter tuning and processing are carried out;
and 2.5, the output data of the step 2.4 is connected into a Softmax regression classifier, and the classifier is used for converting the characteristic value into 0 or 1 output which can be used for classification, so that the function of trafficability prediction is realized.
As a further technical solution of the present invention, the LSTM unit adopted in the substep 2.1 of the second step may determine whether the input information is useful according to a rule, and the core is to control the state c of the unit with a 'gate'. At the current time t, forget the door f t C responsible for controlling the last moment t-1 How much information is saved to c at the current time t The method comprises the steps of carrying out a first treatment on the surface of the Input gate i t Selecting instant state c at the current time t * How much information is input to the current cell; output door o t Control of current state c t How much information is output h as hidden layer at that time t The method comprises the steps of carrying out a first treatment on the surface of the Instant alternative state c t * From input x of network at current time t t And hidden layer output h at last moment t-1 Together, they are calculated by the following formulas
Wherein W and b are weight matrix and bias term, respectively, subscripts thereoff、i、o、cRepresenting the states of forget gate, input gate, output gate and alternative states respectively;for the activation function, the activation of the 'gate' state uses the sigmoid function +.>Instant status c t * Is activated by the tanh function->. Finally, hidden layer output h of LSTM unit at current moment t From the output gate o t And the current cell state c t Determining together; current state c t The output of (a) can be obtained by forgetting the gate f t State c at last moment t-1 Input gate i t Instant status c t * Determining together; while the LSTM cell outputs +_ in the instant state at the next instant>Can be outputted by hidden layer h t The weight matrix W of the output layer and the bias term b are determined together, and the calculation formulas of the weight matrix W and the bias term b are respectively as follows
In the method, in the process of the invention,the symbolic representation is multiplied by the element.
The feature value obtained after the fusion of the cascade layers in the substep 2.3 of the second step is expressed as a one-dimensional feature vector:
in the method, in the process of the invention,respectively represent a traction coefficient characteristic value, a driving efficiency characteristic, and a longitudinal speed characteristic corresponding to the ith sample, which are scalar values.
As a further aspect of the present invention, the outputs of the full-connection layer in substep 2.4 of the second step are mutually independent at each time t, and at each time t, the outputs of the full-connection layer can be expressed as followsIn the form of (1), wherein X t Input from the previous layer.
As a further technical scheme of the invention, the probability estimation formula of calculating the j-th category by the Softmax regression classifier in the substep 2.5 of the second step is as follows
In the method, in the process of the invention,the matrix composed of two category parameters is represented, and the value of the category j is 1 or 2. At this time, the liquid crystal display device,input eigenvalue representing the ith sample from the upper layer network,/for example>For the label value of whether the robot passes or not under the ith sample, they can be collectively denoted as the sample pair +.>。
As a further technical scheme of the invention, in the LSTM network training of the second step and the third step, the invention adopts a loss function for calculation, wherein the loss function is a cross entropy cost function matched with a Softmax regression classifier, and the calculation formula is as follows:
where N is the number of samples,truth value indicating that the ith sample belongs to the jth class,/->The predicted value obtained for the Softmax function classifier.
As a further technical solution of the present invention, the model parameter adjustment in the third step has the following features: firstly, a momentum random gradient descent method is selected as a model solving method; secondly, setting the initial learning rate to be 0.1, and setting the initial maximum iteration period to be 50; and thirdly, a ten-fold cross verification mode is selected for adjustment.
The mobile robot disclosed by the invention is a wheeled mobile robot which is widely applied to field operation or running on soft soil (muddy, mud flat, sand and snow), such as a common electric trolley, a star detection vehicle, a desert detection platform, an unmanned off-road vehicle and the like, and can be applied to front wheel drive, four-wheel drive or wheeled crawler robots.
Compared with the prior art, the invention has the beneficial effects that:
the invention provides a soft ground trafficability prediction method of a mobile robot based on an LSTM network, which can predict whether the robot can smoothly pass through a soft ground in advance under the condition of comprehensively considering the dynamic property, flexibility and economy of the robot, namely, the problems of slipping, sinking and the like of the soft ground are avoided. The trafficability of the traditional moving mechanism is mostly measured by adopting a single evaluation index, and the single evaluation index does not have the functions of estimation and judgment.
Drawings
Fig. 1 is an algorithmic framework and principles of the present invention.
FIG. 2 is a model of LSTM network designed in step two of the present invention.
Fig. 3 is a block diagram of an LSTM network element of the present invention.
Fig. 4 is a schematic diagram of data collection of the mobile robot of the present invention on different floors.
Fig. 5 shows examples of the trafficability model prediction results of the mobile robot of the present invention on three kinds of soft ground.
Detailed Description
The invention is further elucidated below in connection with the drawings and the detailed description, without however limiting the scope of the invention to the embodiments described.
As shown in fig. 1, a method for predicting the trafficability of a soft ground of a mobile robot based on an LSTM network includes the following steps:
step one: three types of trafficability indicating data of the mobile robot are measured and recorded in real time, wherein the three types of trafficability indicating data comprise traction coefficient, driving efficiency and longitudinal speed, and the traction coefficient, the driving efficiency and the longitudinal speed are used as input data of a model; meanwhile, the ground passing condition under the current working condition of artificial observation is used as a tag of input data.
Step two: constructing a soft ground trafficability network prediction model facing the mobile robot based on the LSTM recurrent neural network unit; and sending a large amount of tagged data obtained in the first step into the model for training.
Step three: adjusting the model parameters in the second step and training for multiple times until a stable convergent passing prediction model is obtained; the model can give corresponding pass degree predicted values according to the three indexes which are newly input.
The measurement or calculation method of the three trafficability indexes in the first step is as follows: the traction coefficient pi is defined as the traction force of the hook per unit weight of the robot, i.eWherein D and F Z The traction force and the vertical load of the hook of the mobile robot are respectively measured by a wheel force sensor arranged on a driving wheel of the robot; the driving efficiency e is defined as the ratio of the output power to the input power of the motor on the driving wheel, i.e. +.>Wherein the driving torque T is directly measured by the wheel force sensor, the wheel angular speed +.>The output of the wheel force sensor is obtained through differential calculation by a rotary encoder arranged in the wheel force sensor; the horizontal forward speed of the mobile robot is the longitudinal speed, and can be measured by inertia arranged on the center of the chassis of the robotThe measuring unit gives in real time.
In a preferred embodiment of the present invention, the specific sensor data transmission and reading may be implemented as follows: the micro control unit of the wheel force sensor can pack measured traction force of the hook, vertical load and angle output data of the encoder together, and send the data to the upper computer in a Bluetooth mode; the output data of the inertial measurement unit is directly communicated with the upper computer through the USB serial port of the inertial measurement unit. The LabVIEW software installed on the upper computer can read the data at each moment and adds a waveform icon module in the background, so that the front panel of the LabVIEW software can display a numerical waveform chart.
The data collection of the three passing indexes in the first step should take the same time window and in one time windowTThe data in the model is taken as an actual sampling sample, the sampling frequency is set to be 1Hz, and the number of samples under the current ground working condition is at least 2000, so that the accuracy of the model is ensured; the trafficability of the soft ground of the robot observed by people adopts a form of 0 or 1 to code, namely the robot can be marked as 1 through the ground, and the robot can not pass through (such as sinking and slipping) and is marked as 0.
In a preferred embodiment of the present invention, in order to achieve a preferred effect, the data format described by the method of the present invention may be set as follows: the time window T of the three trafficability indexes is set to be 1min, and the input data of the ith sample at this time can be respectively expressed as
,/>The total number of samples collected was 4000, 80% of which was used for training of the network model, and the remaining 20% was used for test verification.
The pass prediction model in the second step adopts a deep learning algorithm framework based on an LSTM module, and the algorithm framework is designed according to the following form, as shown in fig. 2:
step 2.1, preprocessing the trafficability index data obtained in the step one, and accessing the data of each time t in each sample into an LSTM unit, wherein each LSTM unit is a standard recurrent neural network module comprising a forgetting gate, an input gate and an output gate;
step 2.2, designing a large number of LSTM units in the step 2.1 into a double-layer structure type LSTM network, wherein the first layer LSTM network is used for primary feature extraction, the second layer is used for deep effective feature extraction, and three pieces of trafficability index data respectively correspond to three groups of LSTM networks with the same structure;
step 2.3, the characteristic data of each index obtained in the step 2.2 are accessed into a cascade layer, and the characteristic values of the respective indexes are rearranged and integrated by the cascade layer to obtain a fused characteristic vector;
step 2.4, the feature vector obtained in the step 2.3 is accessed into a full connection layer, and further parameter tuning and processing are carried out;
and 2.5, the output data of the step 2.4 is connected into a Softmax regression classifier, and the classifier is used for converting the characteristic value into 0 or 1 output which can be used for classification, so that the function of trafficability prediction is realized.
As shown in fig. 3, each LSTM cell can determine whether the input information is useful according to a rule, and the core is to control the state of the cell with 'gates'. At the current time t, forget the door f t C responsible for controlling the last moment t-1 How much information is saved to c at the current time t The method comprises the steps of carrying out a first treatment on the surface of the Input gate i t Selecting instant state c at the current time t * How much information is input to the current cell; output door o t Control of current state c t How much information is output h as hidden layer at that time t The method comprises the steps of carrying out a first treatment on the surface of the Instant alternative state c t * From input x of network at current time t t And hidden layer output h at last moment t-1 Together, they are calculated by the following formulas
Wherein W andb is a weight matrix and a bias term, respectively, the subscripts thereoff、i、o、cRepresenting the states of forget gate, input gate, output gate and alternative states respectively;for the activation function, the activation of the 'gate' state uses the sigmoid function +.>Instant status c t * Is activated by the tanh function->. Finally, hidden layer output h of LSTM unit at current moment t From the output gate o t And the current cell state c t Determining together; current state c t The output of (a) can be obtained by forgetting the gate f t State c at last moment t-1 Input gate i t Instant status c t * Determining together; while the LSTM cell outputs +_ in the instant state at the next instant>Can be outputted by hidden layer h t The weight matrix W of the output layer and the bias term b are determined together, and the calculation formulas of the weight matrix W and the bias term b are respectively as follows
In the method, in the process of the invention,the symbolic representation is multiplied by the element.
The feature value obtained after the fusion of the cascade layers in the substep 2.3 of the second step is expressed as a one-dimensional feature vector:
in the method, in the process of the invention,respectively represent a traction coefficient characteristic value, a driving efficiency characteristic, and a longitudinal speed characteristic corresponding to the ith sample, which are scalar values.
The outputs of the fully connected layers in substep 2.4 of said step two are mutually independent at each instant t, at each instant t the outputs of the fully connected layers can also be expressed asIn the form of (1), wherein X t Input from the previous layer.
The Softmax function in substep 2.5 of the second step calculates the probability estimation formula of the first category as
In the method, in the process of the invention,the matrix composed of two category parameters is represented, and the value of the category j is 1 or 2. At this time, the liquid crystal display device,input eigenvalue representing the ith sample from the upper layer network,/for example>For the label value of whether the robot passes or not under the ith sample, they can be collectively denoted as the sample pair +.>。
In the LSTM network training of the second step and the third step, the loss function adopted by the invention is a cross entropy cost function matched with a Softmax regression classifier, and the calculation formula is as follows:
where N is the number of samples,truth value indicating that the ith sample belongs to the jth class,/->The predicted value obtained for the Softmax function classifier.
The model parameter adjustment in the third step has the following characteristics: firstly, a momentum random gradient descent method is selected as a model solving method; secondly, setting the initial learning rate to be 0.1, and setting the initial maximum iteration period to be 50; and thirdly, a ten-fold cross verification mode is selected for adjustment.
In a preferred embodiment of the present invention, the network algorithm model architecture of the present invention may implement LSTM network construction based on a TensorFlow environment, and the super parameter adjustment setting includes: the input time step 60, the input feature dimension is 3, the momentum random gradient adopts an Adam optimizer, the LSTM unit input layer node number is 100, the LSTM unit 32 and the Softmax layer node number is 3, and the adopted portable hardware platform is provided with a Central Processing Unit (CPU) of Intel (R) Core (TM) i7-7700, the main frequency is 3.60 GHz and the cache RAM 16.0 GB, so that the training requirement can be met.
As shown in FIG. 4, the invention selects three different plots to test the effectiveness of the method, including dry sand, grasslands and mud, and the travelling path of the mobile robot is better along an approximate straight line because different types of floors have larger characteristic differences; in order to improve the actual prediction effect, the mobile robot can acquire the passing data in the same land block, and the area of a training data acquisition area under the same land block is 3 times that of an actual test area. FIG. 5 shows the result of the trafficability prediction of the actual test areas of three different plots; from the results, the model gives a higher accuracy of real-time trafficability prediction when the mobile robot runs on the ground.
The foregoing is an example of the present invention and is not intended to limit the invention. All equivalents and alternatives falling within the spirit of the invention are intended to be included within the scope of the invention. What is not elaborated on the invention belongs to the prior art which is known to the person skilled in the art.
Claims (10)
1. The mobile robot soft ground trafficability prediction method based on the LSTM network is characterized by comprising the following steps of:
step one: three types of trafficability indicating data of the mobile robot are measured and recorded in real time, wherein the three types of trafficability indicating data comprise traction coefficient, driving efficiency and longitudinal speed, and the traction coefficient, the driving efficiency and the longitudinal speed are used as input data of a model; meanwhile, the ground passing condition under the current working condition of artificial observation is used as a label of input data;
step two: based on the LSTM unit, constructing a soft ground trafficability network prediction model facing the mobile robot, and sending a large amount of labeled data obtained in the step one into the model for training;
step three: adjusting the model parameters in the second step and training for multiple times until a stable convergent passing prediction model is obtained; the model can give corresponding pass degree predicted values according to the three indexes which are newly input.
2. The LSTM network-based mobile robot soft ground trafficability prediction method according to claim 1, wherein the three trafficability indexes in the first step are measured or calculated as follows: the traction coefficient pi is defined as the traction force of the hook per unit weight of the robot, i.eWherein D and F Z The traction force and the vertical load of the hook of the mobile robot are respectively; the driving efficiency e is defined as the ratio of the output power to the input power of the motor on the driving wheel, i.e. +.>The method comprises the steps of carrying out a first treatment on the surface of the The horizontal forward speed V of the mobile robot is the longitudinal speed.
3. The LSTM network-based mobile robot soft ground trafficability prediction method of claim 2, wherein D and F Z Measured by a wheel force sensor arranged on a driving wheel of the robot; the driving torque T is directly measured by a wheel force sensor, and the angular velocity of the wheelThe longitudinal speed is obtained by differential calculation of the output obtained by a rotary encoder built in a wheel force sensorvIs given in real time by an inertial measurement unit arranged on the central position of the chassis of the robot.
4. A method for predicting the trafficability of a soft ground of a mobile robot based on an LSTM network according to claim 2 or 3, wherein the data collection of the three trafficability indexes in the first step has the following features: the same time window is adopted for data acquisition, and the data acquisition is carried out in one time windowTThe data in the sample is defined as one sample, and the sampling frequency is 1Hz; the pass degree label of the human observation adopts a 0 or 1 form code, namely, the robot can be marked as 1 through the ground, and the robot can not pass through the ground and is marked as 0.
5. The method for predicting the trafficability of the soft ground of the mobile robot based on the LSTM network according to claim 4, wherein the trafficability network prediction model in the second step adopts a deep learning algorithm framework based on the LSTM module, and comprises the following specific steps:
step 2.1, preprocessing the trafficability index data obtained in the step one, and accessing the data of each time t in each sample into an LSTM unit, wherein each LSTM unit is a standard recurrent neural network module comprising a forgetting gate, an input gate and an output gate;
step 2.2, designing a large number of LSTM units in the step 2.1 into a double-layer structure type LSTM network, wherein the first layer LSTM network is used for primary feature extraction, the second layer is used for deep effective feature extraction, and three pieces of trafficability index data respectively correspond to three groups of LSTM networks with the same structure;
step 2.3, the characteristic data of each index obtained in the step 2.2 are accessed into a cascade layer, and the characteristic values of the respective indexes are rearranged and integrated by the cascade layer to obtain a fused characteristic vector;
step 2.4, the feature vector obtained in the step 2.3 is accessed into a full connection layer, and further parameter tuning and processing are carried out;
and 2.5, the output data of the step 2.4 is connected into a Softmax regression classifier, and the classifier is used for converting the characteristic value into 0 or 1 output which can be used for classification, so that the function of trafficability prediction is realized.
6. The LSTM network-based mobile robot soft ground trafficability prediction method according to claim 5, wherein the feature values of the cascade layer fusion in the substep 2.3 are expressed as one-dimensional feature vectors:the method comprises the steps of carrying out a first treatment on the surface of the In (1) the->Respectively represent a traction coefficient characteristic value, a driving efficiency characteristic, and a longitudinal speed characteristic corresponding to the ith sample, which are scalar values.
7. The LSTM network-based mobile robot soft floor trafficability prediction method of claim 5, wherein the outputs of the full link layers in the substep 2.4 are independent of each other at each time t, and the outputs of the full link layers at each time t are also expressed asIn the form of (1), wherein X t From the formerLayer input.
8. The LSTM network-based mobile robot soft ground trafficability prediction method of claim 5, wherein the Softmax regression classifier in sub-step 2.5 computes a probability estimate for the j-th class expressed as:
the method comprises the steps of carrying out a first treatment on the surface of the In (1) the->Representing a matrix of two class parameters, class j having a value of 1 or 2,/->For the input feature of the ith sample from the upper layer network, +.>A tag value indicating whether or not the robot of the i-th sample passes.
9. The LSTM network-based mobile robot soft ground trafficability prediction method according to claim 1 or 2, wherein in the training of the second and third steps, a loss function is used for calculation, where the loss function is a cross entropy cost function matched with a Softmax regression classifier, and is expressed as follows:the method comprises the steps of carrying out a first treatment on the surface of the Where N is the number of samples,truth value indicating that the ith sample belongs to the jth class,/->Is SoftmaxAnd a predicted value obtained by the function classifier.
10. The LSTM network-based mobile robot soft ground trafficability prediction method according to claim 1 or 2, wherein the model parameter adjustment in the third step has the following features: firstly, a momentum random gradient descent method is selected as a model solving method; secondly, setting the initial learning rate to be 0.1, and setting the initial maximum iteration period to be 50; and thirdly, a ten-fold cross verification mode is selected for 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 CN109583571A (en) | 2019-04-05 |
CN109583571B true 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) |
Families Citing this family (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110390386B (en) * | 2019-06-28 | 2022-07-29 | 南京信息工程大学 | Sensitive long-short term memory method based on input change differential |
CN110362081B (en) * | 2019-07-16 | 2020-04-24 | 武昌理工学院 | Mobile robot path planning method |
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 |
CN111679970B (en) * | 2020-05-14 | 2022-09-27 | 南京理工大学 | Method for predicting running environment state of robot software system |
CN111651935B (en) * | 2020-05-25 | 2023-04-18 | 成都千嘉科技股份有限公司 | Multi-dimensional expansion prediction method and device for non-stationary time series data |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
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 |
Family Cites Families (1)
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 |
-
2018
- 2018-12-05 CN CN201811480067.4A patent/CN109583571B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
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 |
---|
车辆地面通过性测试***设计;魏凤;《中国优秀硕士学位论文全文数据库工程科技Ⅱ辑》;20170315;第5-52页 * |
Also Published As
Publication number | Publication date |
---|---|
CN109583571A (en) | 2019-04-05 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109583571B (en) | Mobile robot soft ground trafficability prediction method based on LSTM network | |
Sadrossadat et al. | Prediction of the resilient modulus of flexible pavement subgrade soils using adaptive neuro-fuzzy inference systems | |
Zou et al. | Identification of the shear parameters for lunar regolith based on a GA-BP neural network | |
Gonzalez et al. | Characterization of machine learning algorithms for slippage estimation in planetary exploration rovers | |
Duchanoy et al. | A novel recurrent neural network soft sensor via a differential evolution training algorithm for the tire contact patch | |
CN106023344A (en) | Driving style estimation method based on driving pattern transition probability | |
Bouguelia et al. | Unsupervised classification of slip events for planetary exploration rovers | |
Reina et al. | Slip-based terrain estimation with a skid-steer vehicle | |
CN111598142B (en) | Outdoor terrain classification method for wheeled mobile robot | |
CN107526886B (en) | Method for determining lunar soil mechanical parameters in lunar surface actual walking process | |
Sebastian et al. | Support vector machine based real-time terrain estimation for tracked robots | |
Skonieczny et al. | Data‐driven mobility risk prediction for planetary rovers | |
Goodin et al. | Fast terrain traversability estimation with terrestrial lidar in off-road autonomous navigation | |
Gopalakrishnan | Effect of training algorithms on neural networks aided pavement diagnosis | |
Heyns et al. | A method for real-time condition monitoring of haul roads based on Bayesian parameter estimation | |
Gharieb et al. | Modeling of pavement roughness utilizing artificial neural network approach for Laos national road network | |
CN108510130A (en) | A kind of Intelligent unattended drives vehicle limit mileage intelligence multi-source appraisal procedure and device | |
Zhang et al. | AI-based modeling and data-driven identification of moving load on continuous beams | |
Zhang et al. | Direction-decision learning based pedestrian flow behavior investigation | |
Vosahlik et al. | Self-supervised learning of camera-based drivable surface friction | |
Nourizadeh et al. | In situ slip estimation for mobile robots in outdoor environments | |
Pan et al. | Maneuver-based deep learning parameter identification of vehicle suspensions subjected to performance degradation | |
Fu et al. | Novel hybrid physics-informed deep neural network for dynamic load prediction of electric cable shovel | |
Li et al. | Efficient method based on recurrent neural networks for pavement evenness detection | |
CN108656992A (en) | Automatic driving vehicle power supply wisdom prediction technique and device under a kind of Severe rainstorm environment |
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 |