WO2023035462A1 - 基于脉冲神经网络的小车轮速自调控方法及自调控系统 - Google Patents

基于脉冲神经网络的小车轮速自调控方法及自调控系统 Download PDF

Info

Publication number
WO2023035462A1
WO2023035462A1 PCT/CN2021/137998 CN2021137998W WO2023035462A1 WO 2023035462 A1 WO2023035462 A1 WO 2023035462A1 CN 2021137998 W CN2021137998 W CN 2021137998W WO 2023035462 A1 WO2023035462 A1 WO 2023035462A1
Authority
WO
WIPO (PCT)
Prior art keywords
neuron
value
pwm
wheel speed
trolley
Prior art date
Application number
PCT/CN2021/137998
Other languages
English (en)
French (fr)
Inventor
高钢
李骁健
王成
Original Assignee
中国科学院深圳先进技术研究院
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 中国科学院深圳先进技术研究院 filed Critical 中国科学院深圳先进技术研究院
Publication of WO2023035462A1 publication Critical patent/WO2023035462A1/zh

Links

Images

Classifications

    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B60—VEHICLES IN GENERAL
    • B60L—PROPULSION OF ELECTRICALLY-PROPELLED VEHICLES; SUPPLYING ELECTRIC POWER FOR AUXILIARY EQUIPMENT OF ELECTRICALLY-PROPELLED VEHICLES; ELECTRODYNAMIC BRAKE SYSTEMS FOR VEHICLES IN GENERAL; MAGNETIC SUSPENSION OR LEVITATION FOR VEHICLES; MONITORING OPERATING VARIABLES OF ELECTRICALLY-PROPELLED VEHICLES; ELECTRIC SAFETY DEVICES FOR ELECTRICALLY-PROPELLED VEHICLES
    • B60L15/00—Methods, circuits, or devices for controlling the traction-motor speed of electrically-propelled vehicles
    • B60L15/20—Methods, circuits, or devices for controlling the traction-motor speed of electrically-propelled vehicles for control of the vehicle or its driving motor to achieve a desired performance, e.g. speed, torque, programmed variation of speed
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B62—LAND VEHICLES FOR TRAVELLING OTHERWISE THAN ON RAILS
    • B62D—MOTOR VEHICLES; TRAILERS
    • B62D11/00—Steering non-deflectable wheels; Steering endless tracks or the like
    • B62D11/02—Steering non-deflectable wheels; Steering endless tracks or the like by differentially driving ground-engaging elements on opposite vehicle sides
    • B62D11/04—Steering non-deflectable wheels; Steering endless tracks or the like by differentially driving ground-engaging elements on opposite vehicle sides by means of separate power sources
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05D—SYSTEMS FOR CONTROLLING OR REGULATING NON-ELECTRIC VARIABLES
    • G05D3/00—Control of position or direction
    • G05D3/12—Control of position or direction using feedback
    • G05D3/20—Control of position or direction using feedback using a digital comparing device
    • 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
    • G06N3/08—Learning methods
    • G06N3/09—Supervised learning
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/60—Other road transportation technologies with climate change mitigation effect
    • Y02T10/72—Electric energy management in electromobility

Definitions

  • the invention relates to the field of intelligent trolley control, in particular to a method and system for self-regulating the speed of trolleys based on pulse neural networks.
  • the smart car is especially suitable for working in harsh environments, and the car has the characteristics of convenient control and simple algorithm deployment and testing. This technology can be widely used in unmanned vehicles, unmanned production lines, warehouses, service robots and other fields. In the field of education, the actual teaching of intelligent algorithms can also be carried out through smart cars.
  • differential drive In the current smart car, differential drive is still the mainstream.
  • the so-called “differential drive” means that the left and right wheels are driven by motors respectively, and the car moves forward and turns by changing the speed of the two motors.
  • using this method for control will cause the car to be unable to "walk in a straight line” stably.
  • Open-loop control without a feedback mechanism cannot eliminate the above-mentioned speed error of the left and right wheels, because the above-mentioned disturbance is random. If the trolley wants to go in a straight line, the only way to achieve closed-loop control is to give timely feedback to the left and right wheels when the trolley is disturbed, and to correct the speed deviation of the two wheels, so that it can walk out of a straight line.
  • the combination of PID algorithm and wheel speed detection device is a common method used to solve the problem of "not going straight" of the car, but it is not suitable for complex environments and the adaptive ability is weak, and the wheel speed detection device is not suitable for off-the-shelf
  • the trolley motor scheme is not necessarily suitable.
  • the present invention provides a method and system for self-regulating the speed of trolley wheels based on the impulse neural network.
  • the specific plan is as follows:
  • a method for self-regulating the speed of trolley wheels based on a pulse neural network comprising the steps of:
  • adjusting the wheel speed value through the supervised learning algorithm of the spiking neural network combined with the angle signal specifically includes:
  • the first value adjustment includes scaling the first PWM value and the second PWM value, and linearly mapping to the [0, 1] interval;
  • the second value adjustment includes scaling the offset, which is linearly mapped to the interval [-1, 1].
  • the expression of the first value adjustment is:
  • x' represents the PWM adjustment signal
  • x represents the first PWM value or the second PWM value
  • min(x) represents the minimum value of the first PWM value or the minimum value of the second PWM value value
  • max(x) represents the maximum value of the first PWM value or the maximum value of the second PWM value
  • ⁇ ' represents the angle signal
  • ⁇ represents the offset
  • max ( ⁇ ) represents the maximum value of the offset
  • min ( ⁇ ) represents the minimum value of the offset
  • mean ( ⁇ ) represents the average value of the offset value.
  • the abstract expression of the neuron activity is:
  • a represents the activity of the neuron
  • ⁇ is the scaling factor related to the neuron
  • e represents the encoder of the neuron
  • x represents the information to be encoded
  • G[ ] represents the non-linearity determined by the preset neuron model.
  • the decoder expressions include:
  • d represents the decoder
  • a i represents the activity of neuron i
  • a j represents the activity of neuron j
  • T and r are intermediate variables
  • x represents the input data.
  • the supervised learning algorithm is a Prescribed Error Sensitivity supervised learning algorithm
  • the obtained weight correction value ⁇ ij can be obtained by the following formula:
  • ⁇ ij represents the weight correction value
  • E represents the error vector that you want to minimize
  • ⁇ represents the learning rate
  • d represents the decoder
  • n is the dimension of the decoder
  • ⁇ represents the connection weight
  • ⁇ represents the scaling factor related to neurons
  • e represents the encoder of the neuron
  • i, j represent the neuron
  • a i represents the activity of the neuron i.
  • the corrected PWM value is obtained by integrating the neuron activity, the neuron decoder and the connection weight, and the specific expression is:
  • a represents the neuron activity
  • ⁇ represents the connection weight between neurons
  • d represents the neuron decoder
  • P represents the corrected PWM value
  • a self-regulating system of trolley wheel speed based on pulse neural network including the following:
  • Wheel speed acquisition unit used to acquire the first PWM value of the left wheel motor of the trolley and the second PWM value of the right wheel motor of the trolley;
  • Wheel speed adjustment unit with a built-in control state processor, which is used to adjust the first value of the first PWM value and the second PWM value through the control state processor to obtain a neural network that can adapt to the preset pulse The PWM control signal;
  • Offset acquisition unit used to acquire the offset of the motion angle of the trolley recorded by the inertial measurement unit;
  • Offset adjustment unit a built-in angle processor, which is used to adjust the second value through the angle processor to obtain an angle signal that can adapt to the pulse neural network;
  • Adjustment control unit a built-in pulse neural network, which is used to adjust the PWM control signal through the supervised learning algorithm of the pulse neural network, combined with the angle signal, to obtain the left PWM value and the right PWM value;
  • Speed application unit used to apply the left PWM value and the right PWM value to the left wheel motor and the right wheel motor of the trolley respectively to adjust the trolley wheel speed;
  • Real-time adjustment unit used to cycle through the wheel speed acquisition unit, the wheel speed adjustment unit, the offset acquisition unit, the offset adjustment unit, the adjustment control unit, and the speed application unit in order to realize Real-time adjustment of trolley wheel speed.
  • the adjustment control unit specifically includes:
  • Neuron activity module based on a preset neuron model, using the PWM control signal output by the control state processor as a training signal, combined with the nonlinear neural activation function of the neuron model, to obtain an abstract neuronal activity;
  • Decoder module used to obtain a neuron decoder based on the neuron activity and the PWM control signal;
  • Connection weight module used to use the angle signal output by the angle processor as a learning signal, use the supervised learning algorithm of the spiking neural network as a learning algorithm, combine the neuron decoder and the neuron activity to obtain Connection weights between neurons;
  • An output module used to synthesize the neuron activity, the neuron decoder and the connection weight to obtain the corrected left PWM value and the right PWM value.
  • the wheel speed adjustment unit includes scaling the first PWM value and the second PWM value, and linearly mapping to the interval [0, 1];
  • the offset adjustment unit includes scaling the offset and linearly mapping it to the interval [-1, 1].
  • the invention provides a self-regulating method and a self-regulating system of trolley speed based on pulse neural network.
  • the self-regulation method of trolley wheel speed utilizes the characteristics of low power consumption and high dynamics of the pulse neural network, and combines the pulse neural network and IMU to control the trolley going straight, so that the trolley has the ability to adapt to the wheel speed, and solves the problem of the trolley due to its own The problem of "can't walk straight" caused by factors such as different motor characteristics and external environmental interference.
  • the scheme provided by the present invention can realize the straight running of the trolley by adjusting the wheel speed in real time even in a complex environment, has a high degree of adaptability and anti-interference ability, and can The types of compatible smart cars are also wider.
  • Fig. 1 is the schematic flow chart of the trolley wheel speed self-regulating method of the embodiment of the present invention
  • Fig. 2 is a schematic diagram of the principle of the trolley wheel speed self-regulating method according to the embodiment of the present invention
  • Fig. 3 is a schematic diagram of the principle formula of the trolley wheel speed self-regulating method according to the embodiment of the present invention.
  • FIG. 4 is a schematic flow chart of step 105 in an embodiment of the present invention.
  • Fig. 5 is the experimental result figure of the embodiment of the present invention.
  • Fig. 6 is a structural block diagram of the trolley wheel speed self-regulating system of the embodiment of the present invention.
  • Fig. 7 is a schematic structural diagram of an adjustment control unit according to an embodiment of the present invention.
  • This embodiment proposes a self-regulating method of trolley wheel speed based on the pulse neural network, and realizes the straight-going control of the trolley through the combination of the pulse neural network and the IMU, which solves the problem caused by the different characteristics of the trolley itself and the interference of the external environment. "Can't walk straight" problem.
  • the schematic flow chart of the self-regulating method is shown in Figure 1 of the description. The specific plan is as follows:
  • a method for self-regulating the speed of trolley wheels based on a pulse neural network comprising the steps of:
  • the PWM control signal is adjusted to obtain the left PWM value and the right PWM value;
  • the wheel speed control of the trolley adopts pulse width modulation (PWM).
  • Pulse Width Modulation is a very effective technology that uses the digital output of the microprocessor to control the analog circuit. It obtains the required waveform equivalently by modulating the pulse width, so as to realize the speed control of the trolley. control.
  • 101 Acquire the first PWM value of the left wheel motor of the trolley and the second PWM value of the right wheel motor of the trolley.
  • the left wheel and the right wheel of the trolley are each driven by a motor, the left wheel motor drives the left wheel of the trolley, and the right wheel motor drives the right wheel of the trolley.
  • the PWM value of the car changes with the speed of the car. Due to the limitations of the manufacturing process, the driving characteristics of the motors that control the two wheels cannot be completely the same, the shape and size of the two motors cannot be completely the same, and the accuracy of assembly will also vary; in addition, when the car is running, The tires may slip or encounter small obstacles when rolling, which will cause the speed difference between the left and right wheels, so that you can't walk straight.
  • the stable straight running of the trolley is realized by collecting and controlling the PWM values of the left wheel motor and the right wheel motor of the trolley in real time.
  • 102 Perform a first numerical adjustment on the first PWM value and the second PWM value through a preset control state processor to obtain a PWM control signal capable of adapting to a preset pulse neural network. Since the PWM value of the car varies, in order to adapt the PWM value to the neuron model of the pulse neural network, its value needs to be scaled.
  • the scaling manner adopted for the first value adjustment includes but is not limited to any known numerical scaling manner. This embodiment takes rescaling as an example.
  • the first numerical adjustment is performed on the first PWM value and the second PWM value by controlling the state processor, and they are linearly mapped to the interval [0,1], so as to adapt to the neuron model of the spiking neural network.
  • the expression of the first value adjustment is:
  • x' represents the adjusted output signal, which is the PWM adjusted signal in this embodiment.
  • x represents the input signal before adjustment, which is the first PWM value or the second PWM value in this embodiment.
  • min(x) represents the minimum value of the input signal, which is the minimum value of the first PWM value or the minimum value of the second PWM value in this embodiment.
  • max(x) represents the maximum value of the input signal, which is the maximum value of the first PWM value or the maximum value of the second PWM value in this embodiment.
  • 103 Obtain the offset of the movement angle of the trolley recorded by the preset inertial measurement unit.
  • an inertial measurement unit is used to measure the offset of the trolley when it moves.
  • An inertial measurement unit is a device that measures the three-axis attitude angle (or angular rate) and acceleration of an object.
  • Gyroscopes and accelerometers are the main components of the IMU, and their accuracy directly affects the accuracy of the inertial system.
  • the gyroscope and accelerometer produce errors. From the initial alignment, the navigation error increases with time, especially the position error. This is the inertial navigation system. major drawback. Therefore, it is necessary to use external information for assistance to realize integrated navigation, so as to effectively reduce the problem of error accumulation over time. To increase reliability, it is also possible to equip more sensors per axis.
  • the IMU should be installed on the center of gravity of the measured object. Compared with the traditional wheel speed detection device, the inertial measurement unit has a wider adaptability to the trolley model and is easier to install on the trolley.
  • 104 perform a second numerical adjustment on the offset by a preset angle processor to obtain an angle signal that can be adapted to the spiking neural network.
  • the processing of the offset is similar to the processing of the PWM value, both of which are to adapt to the neuron model of the spiking neural network.
  • the scaling manner adopted for the second numerical adjustment includes but is not limited to any known numerical scaling manner.
  • This embodiment takes rescaling as an example.
  • the second numerical adjustment is performed on the offset by the angle processor, so as to adapt to the neuron model of the spiking neural network.
  • the second value adjustment is to linearly map the offset to the interval [-1,1] to obtain the angle signal, and use the positive and negative values of the angle signal to represent the offset direction of the car.
  • the expression of the second value adjustment is:
  • ⁇ ' represents the adjusted output signal, which is an angle signal in this embodiment.
  • ⁇ represents the input signal before adjustment, which is the offset in this embodiment.
  • max( ⁇ ) represents the maximum value of the input signal, which is the maximum value of the offset in this embodiment,
  • min( ⁇ ) represents the minimum value of the input signal, which is the minimum value of the offset in this embodiment, mean( ⁇ ) represents the average value of the offset.
  • 105 Adjust the PWM control signal through a supervised learning algorithm of the pulse neural network combined with the angle signal to obtain the left PWM value and the right PWM value.
  • This step is realized by the pulse neural network.
  • the pulse neural network takes the PWM control signal output by the control state processor as the training signal, takes the angle signal output by the angle processor as the learning signal, and uses the corresponding pulse neural network supervised learning algorithm as the learning algorithm. Adjust the PWM control signal.
  • Step 105 is shown in Figure 4 of the description, specifically including:
  • the PWM control signal output by the control state processor is used as the training signal, combined with the nonlinear neural activation function of the neuron model, to obtain abstract neuron activity;
  • 10502. Obtain a neuron decoder based on the neuron activity and the PWM control signal
  • Spiking neural network is considered as the third-generation neural network, which is a neural network model with more biological significance and the most similar operating mechanism to the brain.
  • the spiking neural network uses different neuron models for information transmission and calculation. Different neuron models have different characteristics, and a more suitable model can be found by replacing neurons.
  • step 10501 the neuron activity in the spiking neural network can be abstractly expressed as:
  • a represents neuron activity
  • ⁇ is a scale factor related to neuron
  • e represents the encoder of neuron
  • x represents the information to be encoded, which is the PWM control signal.
  • G[ ] represents a nonlinear neural activation function, which is determined by a specific neuron model;
  • step 10502 the decoder d can be calculated according to the following formula:
  • d represents the decoder
  • a i represents the activity of neuron i
  • a j represents the activity of neuron j
  • T and r are intermediate variables
  • x represents the input data, which is the PWM control signal.
  • the supervised learning algorithm is a Prescribed Error Sensitivity supervised learning algorithm.
  • the weight correction value ⁇ ij obtained by using the Prescribed Error Sensitivity supervised learning algorithm can be obtained by the following formula:
  • ⁇ ij represents a weight correction value
  • E represents an error vector to be minimized, which is an angle offset in this embodiment.
  • ⁇ represents the learning rate
  • d represents the decoder
  • n is the dimension of the decoder
  • ⁇ represents the connection weight
  • ⁇ represents the scaling factor associated with the neuron
  • e represents the encoder of the neuron
  • i, j represent the neuron
  • a i represents the activity of neuron i.
  • a represents the neuron activity
  • ⁇ represents the connection weight between neurons
  • d represents the neuron decoder
  • P represents the corrected PWM value
  • the corrected PWM value can be obtained by inputting the PWM control signal into the pulse neural network, that is, the left PWM value and the right PWM value.
  • the pulse neural network has the characteristics of low power consumption and good dynamics.
  • the pulse neural network combined with the inertial measurement unit (IMU) can provide a highly adaptive and relatively low power consumption solution to the problem of the car "not going straight".
  • step 106 apply the left PWM value and the right PWM value to the left wheel motor and the right wheel motor of the trolley respectively to adjust the wheel speed of the trolley.
  • step 105 the correction values of the left and right wheels can be obtained, and the left and right PWM values can be applied to the trolley to adjust the speed of the left and right wheels of the trolley.
  • the accompanying drawing 5 of the description is the experimental result diagram of this embodiment, which is the result of the straight-going control of the trolley after the combination of the pulse neural network and the IMU.
  • the x-axis is time
  • the y-axis is angular offset. It can be clearly seen from the experimental results that after 6s of self-adaptive adjustment, the trolley can walk in a straight line with almost no angular deviation.
  • This embodiment proposes a self-regulating method of trolley wheel speed based on the pulse neural network, which utilizes the characteristics of low power consumption and high dynamics of the pulse neural network, and combines the pulse neural network and the IMU to control the car going straight, so that the trolley It has the ability to adapt to the wheel speed, which solves the problem of "can't walk straight" caused by factors such as its own motor characteristics and external environmental interference.
  • the scheme provided by this embodiment can realize the straight running of the trolley by adjusting the wheel speed in real time even in a complex environment, and has a high degree of adaptability and anti-interference ability, and the The types of smart cars that can be adapted are also more extensive.
  • This embodiment proposes a self-regulating system of trolley speed based on the pulse neural network, adopts a kind of self-regulating method of trolley speed based on the pulse neural network proposed in embodiment 1, the module schematic diagram of the trolley speed self-regulating system is shown in the instruction manual Shown in accompanying drawing 6.
  • the specific plan is as follows:
  • a self-regulating system for trolley wheel speed based on pulse neural network comprising a cycle wheel speed acquisition unit 1, a wheel speed adjustment unit 2, an offset acquisition unit 3, an offset adjustment unit 4, an adjustment control unit 5, and a moving speed application unit 6 and real-time adjustment unit 7 as follows:
  • Wheel speed acquisition unit 1 used to acquire the first PWM value of the left wheel motor of the trolley and the second PWM value of the right wheel motor of the trolley;
  • Wheel speed adjustment unit 2 built-in a control state processor, used to adjust the first PWM value and the second PWM value through the control state processor to obtain a PWM control signal that can adapt to the preset pulse neural network.
  • the wheel speed adjustment unit 2 includes scaling the first PWM value and the second PWM value, and linearly mapping to the interval [0, 1].
  • Offset acquisition unit 3 used to acquire the offset of the movement angle of the trolley recorded by the inertial measurement unit;
  • the offset adjustment unit 4 has a built-in angle processor, which is used to perform a second numerical adjustment on the offset through the angle processor to obtain an angle signal that can be adapted to the spiking neural network.
  • the offset adjustment unit 4 includes scaling the offset, which is linearly mapped to the [-1, 1] interval.
  • Adjustment control unit 5 a built-in pulse neural network, which is used to adjust the PWM control signal through the supervised learning algorithm of the pulse neural network, combined with the angle signal, to obtain the left PWM value and the right PWM value;
  • Speed application unit 6 used to apply the left PWM value and the right PWM value to the left wheel motor and the right wheel motor of the trolley respectively to adjust the trolley wheel speed;
  • Real-time adjustment unit 7 used to cycle wheel speed acquisition unit 1, wheel speed adjustment unit 2, offset acquisition unit 3, offset adjustment unit 4, adjustment control unit 5 and speed application unit 6 in order to realize real-time adjustment of trolley wheel speed Adjustment.
  • the adjustment control unit 5 includes a neuron activity module 51, a decoder module 52, a connection weight module 53 and an output module 54, as shown in Figure 7 of the specification.
  • Neuron activity module 51 based on the preset neuron model, the PWM control signal output by the control state processor is used as the training signal, combined with the nonlinear neural activation function of the neuron model to obtain abstract neuron activity.
  • the abstract expression of the neuron activity in the spiking neural network is:
  • a represents neuron activity
  • ⁇ is a scale factor related to neuron
  • e represents the encoder of neuron
  • x represents the information to be encoded, which is the PWM control signal.
  • G[ ⁇ ] represents the nonlinear neural activation function, which is determined by the specific neuron model.
  • Decoder module 52 used to obtain a neuron decoder based on the neuron activity and the PWM control signal.
  • Decoder d can be calculated according to the following formula:
  • d represents the decoder
  • a i represents the activity of neuron i
  • a j represents the activity of neuron j
  • T and r are intermediate variables
  • x represents the input data, which is the PWM control signal.
  • Connection weight module 53 used to use the angle signal output by the angle processor as the learning signal, use the supervised learning algorithm of the pulse neural network as the learning algorithm, and combine the neuron decoder and neuron activity to obtain the connection weight between neurons .
  • the supervised learning algorithm is a Prescribed Error Sensitivity supervised learning algorithm.
  • the weight correction value ⁇ ij obtained by using the Prescribed Error Sensitivity supervised learning algorithm can be obtained by the following formula:
  • ⁇ ij represents a weight correction value
  • E represents an error vector to be minimized, which is an angle offset in this embodiment.
  • ⁇ represents the learning rate
  • d represents the decoder
  • n is the dimension of the decoder
  • ⁇ represents the connection weight
  • ⁇ represents the scaling factor associated with the neuron
  • e represents the encoder of the neuron
  • i, j represent the neuron
  • a i represents the activity of neuron i.
  • Output module 54 for synthesizing neuron activities, neuron decoders and connection weights to obtain corrected left PWM values and right PWM values.
  • a represents the neuron activity
  • ⁇ represents the connection weight between neurons
  • d represents the neuron decoder
  • P represents the corrected PWM value
  • the corrected PWM value can be obtained by inputting the PWM control signal into the pulse neural network, that is, the left PWM value and the right PWM value.
  • the pulse neural network has the characteristics of low power consumption and good dynamics.
  • the pulse neural network combined with the inertial measurement unit (IMU) can provide a highly adaptive and relatively low power consumption solution to the problem of the car "not going straight".
  • This embodiment proposes a self-regulating system of trolley wheel speed based on a pulse neural network, and systematizes the self-regulating method of trolley wheel speed based on a pulse neural network in Embodiment 1 to make it more practical.
  • the invention provides a self-regulating method and a self-regulating system of trolley speed based on pulse neural network.
  • the self-regulation method of trolley wheel speed utilizes the characteristics of low power consumption and high dynamics of the pulse neural network, and combines the pulse neural network and IMU to control the trolley going straight, so that the trolley has the ability to adapt to the wheel speed, and solves the problem of the trolley due to its own The problem of "can't walk straight" caused by factors such as different motor characteristics and external environmental interference.
  • the scheme provided by the present invention can realize the straight running of the trolley by adjusting the wheel speed in real time even in a complex environment, has a high degree of adaptability and anti-interference ability, and can The types of compatible smart cars are also wider.
  • each module or each step of the present invention described above can be realized by a general-purpose computing device, and they can be concentrated on a single computing device, or distributed on a network formed by multiple computing devices.
  • they can be implemented with executable program codes of computer devices, so that they can be stored in storage devices and executed by computing devices, or they can be made into individual integrated circuit modules, or a plurality of modules in them Or the steps are fabricated into a single integrated circuit module to realize.
  • the present invention is not limited to any specific combination of hardware and software.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Mechanical Engineering (AREA)
  • Transportation (AREA)
  • Combustion & Propulsion (AREA)
  • Chemical & Material Sciences (AREA)
  • Automation & Control Theory (AREA)
  • Power Engineering (AREA)
  • Feedback Control In General (AREA)
  • Electric Propulsion And Braking For Vehicles (AREA)

Abstract

一种基于脉冲神经网络的小车轮速自调控方法，包括：获取小车左轮电机PWM值；通过控制状态处理器对PWM值进行数值调整；获取惯性测量单元记录的小车运动角度的偏移量；通过角度处理器对偏移量进行数值调整；通过脉冲神经网络的有监督学习算法，结合角度信号，对PWM控制信号进行调整并应用到小车上，调整小车轮速；循环上述步骤实现小车轮速的实时调整。利用了脉冲神经网络低功耗、高动态性的特点，将脉冲神经网络和IMU相结合进行小车直走控制，使小车具备轮速自适应能力，解决了小车由于其自身电机特性不相同、外界环境干扰等因素造成的走不直问题。还公开一种小车轮速自调控系统。

Description

基于脉冲神经网络的小车轮速自调控方法及自调控系统 技术领域
本发明涉及智能小车控制领域，具体而言，涉及一种基于脉冲神经网络的小车轮速自调控方法及自调控系统。
背景技术
在险恶的环境下工作，人类必需采取严密的保护措施，而机器人可以进入或穿过这些危险区域进行维护和探测工作，且不需要得到像对人一样的保护。智能小车尤其适用于在条件恶劣的环境中工作，并且小车具有控制方便、算法部署测试简单等特点，该技术可以广泛应用于无人驾驶机动车，无人生产线、仓库、服务机器人等领域。而在教育领域，也可以通过智能小车来进行智能算法的实际教学。
在目前的智能小车中，差分驱动仍是主流。所谓“差分驱动”，就是左、右轮分别用电机驱动，通过改变两个电机的转速实现小车前进和转向。但使用这种方法进行控制会导致小车无法稳定地“走直线”。原因在二：其一、由于制备工艺的限制，控制两个轮子的电机本身的驱动特性不可能完全相同，两个电机的外形大小不可能是完全一致的，组装时的精度也会出现差异；其二、小车在行驶过程中，轮胎在滚动时可能发生打滑或遇到细小的障碍物等情况，这都会造成左右轮的速度出现差异，从而走不直。
没有反馈机制的开环控制是无法消除上述左右轮的速度误差的，因为上述的扰动是随机的。要想小车走一条直线，唯有实现闭环控制，当小车受到扰动时能对左右轮及时给予反馈，修正两轮的速度偏差，从而可以走出一条直线。
目前在智能小车领域，PID算法和轮速检测装置结合是用来解决小车 “走不直”的常用方法，但是其不适用于复杂环境并且自适应能力较弱，且轮速检测装置对于现成的小车电机方案不一定适配。
因此，需要一种能够适用于复杂环境且自适应能力强以使小车走直线的方案，能够解决上述问题。
发明内容
基于现有技术存在的问题，本发明提供了一种基于脉冲神经网络的小车轮速自调控方法及自调控系统。具体方案如下：
一种基于脉冲神经网络的小车轮速自调控方法，包括如下步骤：
获取小车左轮电机的第一PWM值和小车右轮电机的第二PWM值；
通过预设控制状态处理器对所述第一PWM值和所述第二PWM值进行第一数值调整，得到能够适配预设脉冲神经网络的PWM控制信号；
获取预设惯性测量单元记录的小车运动角度的偏移量；
通过预设角度处理器对所述偏移量进行第二数值调整，得到能够适配所述脉冲神经网络的角度信号；
通过所述脉冲神经网络的有监督学习算法，结合所述角度信号，对所述PWM控制信号进行调整，得到左PWM值和右PWM值；
将所述左PWM值和所述右PWM值分别应用到小车的左轮电机和右轮电机上，调整小车轮速；
循环上述步骤实现小车轮速的实时调整。
在一个具体实施例中，“通过所述脉冲神经网络的有监督学习算法，结合所述角度信号，对所述轮速值进行调整”具体包括：
基于预设的神经元模型，以所述控制状态处理器输出的所述PWM控制信号为训练信号，结合所述神经元模型的非线性神经激活函数，得到抽象化的神经元活动；
基于所述神经元活动和所述PWM控制信号，得到神经元解码器；
以所述角度处理器输出的所述角度信号为学习信号，以所述脉冲神经网络的有监督学习算法作为学习算法，结合所述神经元解码器和所述神经元活动，得到神经元之间的连接权重；
综合所述神经元活动、所述神经元解码器和所述连接权重，得到修正后的所述左PWM值和所述右PWM值。
在一个具体实施例中，所述第一数值调整包括对所述第一PWM值和所述第二PWM值进行缩放，线性映射到[0，1]区间；
所述第二数值调整包括对所述偏移量进行缩放，线性映射到[-1，1]区间。
在一个具体实施例中，所述第一数值调整的表达式为：
Figure PCTCN2021137998-appb-000001
其中，x′表示所述PWM调整信号，x表示所述第一PWM值或所述第二PWM值，min(x)表示所述第一PWM值的最小值或所述第二PWM值的最小值，max(x)表示所述第一PWM值的最大值或所述第二PWM值的最大值；
所述第二数值调整的表达式为：
Figure PCTCN2021137998-appb-000002
其中，θ′表示所述角度信号，θ表示偏移量，max(θ)表示偏移量的最大值，min(θ)表示偏移量的最小值，mean(θ)表示偏移量的平均值。
在一个具体实施例中，所述神经元活动的抽象化表达式为：
a＝G[αe·x]
其中，a表示所述神经元活动，α是与神经元相关的标度因子，e表示神经元的编码器，x表示要编码的信息，G[·]表示由预设神经元模型决定的非线性神经激活函数；
所述解码器的表达式包括：
d＝r -1Τ
T ij＝∫a ia jdx
r＝∫a jxdx
其中，d表示解码器，a i表示神经元i的活动，a j表示神经元j的活动，T和r为中间变量，x表示输入数据。
在一个具体实施例中，所述有监督学习算法为Prescribed Error Sensitivity有监督学习算法，得到的权重修正值Δω ij可由以下公式求得：
Figure PCTCN2021137998-appb-000003
其中，Δω ij表示权重修正值，E表示希望最小化的错误向量，κ表示学习率，d表示解码器，n为解码器的维度，ω表示连接权重，α表示与神经元相关的标度因子，e表示神经元的编码器，i、j表示神经元，a i表示神经元i的活动。
在一个具体实施例中，综合所述神经元活动、所述神经元解码器和所述连接权重得到修正后的PWM值，具体表达式为：
Figure PCTCN2021137998-appb-000004
其中，a表示神经元活动，ω表示神经元之间的连接权重，d表示神经元解码器，P表示修正后的PWM值。
一种基于脉冲神经网络的小车轮速自调控系统，包括如下：
轮速获取单元：用于获取小车左轮电机的第一PWM值和小车右轮电机的第二PWM值；
轮速调整单元：内置有控制状态处理器，用于通过所述控制状态处理器对所述第一PWM值和所述第二PWM值进行第一数值调整，得到能够适配预设脉冲神经网络的PWM控制信号；
偏移获取单元：用于获取惯性测量单元记录的小车运动角度的偏移量；
偏移调整单元：内置有角度处理器，用于通过所述角度处理器对进行第二数值调整，得到能够适配所述脉冲神经网络的角度信号；
调整控制单元：内置有脉冲神经网络，用于通过所述脉冲神经网络的有监督学习算法，结合所述角度信号，对所述PWM控制信号进行调整，得到左PWM值和右PWM值；
移速应用单元：用于将所述左PWM值和所述右PWM值分别应用到小车的左轮电机和右轮电机上，调整小车轮速；
实时调整单元：用于依次循环所述轮速获取单元、所述轮速调整单元、所述偏移获取单元、所述偏移调整单元、所述调整控制单元和所述移速应用单元，实现小车轮速的实时调整。
在一个具体实施例中，所述调整控制单元具体包括：
神经元活动模块：用于基于预设的神经元模型，以所述控制状态处理器输出的所述PWM控制信号为训练信号，结合所述神经元模型的非线性神经激活函数，得到抽象化的神经元活动；
解码器模块：用于基于所述神经元活动和所述PWM控制信号，得到神经元解码器；
连接权重模块：用于以所述角度处理器输出的角度信号为学习信号，以所述脉冲神经网络的有监督学习算法作为学习算法，结合所述神经元解码器和所述神经元活动，得到神经元之间的连接权重；
输出模块：用于综合所述神经元活动、所述神经元解码器和所述连接权重，得到修正后的所述左PWM值和所述右PWM值。
在一个具体实施例中，所述轮速调整单元包括对所述第一PWM值和所述第二PWM值进行缩放，线性映射到[0，1]区间；
所述偏移调整单元包括对所述偏移量进行缩放，线性映射到[-1，1]区间。
本发明具有如下有益效果：
本发明提供了一种基于脉冲神经网络的小车轮速自调控方法及自调控系统。小车轮速自调控方法利用了脉冲神经网络低功耗、高动态性的特点，将脉冲神经网络和IMU相结合进行小车直走控制，使小车具备轮速自适应能力，解决了小车由于其自身电机特性不相同、外界环境干扰等因素造成的“走不直”问题。相较于现有PID结合轮速检测装置的方案，本发明提供的方案即使在复杂环境下，也能通过实时调整轮速实现小车直走，具备高度的适应性和抗干扰能力，并且所能适配的智能小车的类型也更为广泛。
为使本发明的上述目的、特征和优点能更明显易懂，下文特举较佳实施例，并配合所附附图，作详细说明如下。
附图说明
为了更清楚地说明本发明实施例的技术方案，下面将对实施例中所需要使用的附图作简单地介绍，应当理解，以下附图仅示出了本发明的某些实施例，因此不应被看作是对范围的限定，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他相关的附图。
图1是本发明实施例的小车轮速自调控方法流程示意图；
图2是本发明实施例的小车轮速自调控方法原理示意图；
图3是本发明实施例的小车轮速自调控方法原理公式示意图；
图4是本发明实施例的步骤105具体流程示意图；
图5是本发明实施例的实验结果图；
图6是本发明实施例的小车轮速自调控系统结构框图；
图7是本发明实施例的调整控制单元结构示意图。
附图标记：1-循环轮速获取单元；2-轮速调整单元；3-偏移获取单元； 4-偏移调整单元；5-调整控制单元；6-移速应用单元；7-实时调整单元；51-神经元活动模块；52-解码器模块；53-连接权重模块；54-输出模块。
具体实施方式
下面将结合本发明实施例中的附图，对本发明实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例仅是本发明的一部分实施例，而不是全部的实施例。基于本发明中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本发明保护的范围。
实施例1
本实施例提出了一种基于脉冲神经网络的小车轮速自调控方法，通过脉冲神经网络和IMU相结合实现小车直走控制，解决了小车由于其自身电机特性不相同、外界环境干扰等因素造成的“走不直”问题。自调控方法的流程示意图如说明书附图1所示。具体方案如下:
一种基于脉冲神经网络的小车轮速自调控方法，包括如下步骤：
101、获取小车左轮电机的第一PWM值和小车右轮电机的第二PWM值；
102、通过预设控制状态处理器对第一PWM值和第二PWM值进行第一数值调整，得到能够适配预设脉冲神经网络的PWM控制信号；
103、获取预设惯性测量单元记录的小车运动角度的偏移量；
104、通过预设角度处理器对所述偏移量进行第二数值调整，得到能够适配脉冲神经网络的角度信号；
105、通过脉冲神经网络的有监督学习算法，结合角度信号，对PWM控制信号进行调整，得到左PWM值和右PWM值；
106、将左PWM值和右PWM值分别应用到小车的左轮电机和右轮电机上，调整小车轮速；
107、循环上述步骤101-106，实现小车轮速的实时调整。
在本实施例中，小车的轮速控制采用脉冲宽度调制(PWM)。脉冲宽度调制是利用微处理器的数字输出来对模拟电路进行控制的一种非常有效的技术，其通过对脉冲的宽度进行调制来等效的获得所需要的波形，以此实现小车轮速的控制。
说明书附图2为本实施例的原理示意图，说明书附图3为说明书附图2中各模块的公式表达示意图。通过闭环控制实现小车轮速的自适应。
具体地，101、获取小车左轮电机的第一PWM值和小车右轮电机的第二PWM值。在本实施例中，小车的左轮和右轮各由一个电机进行驱动，左轮电机驱动小车的左轮，右轮电机驱动小车的右轮。
小车的PWM值的大小是随着小车速度的大小所变化的。由于制备工艺的限制，控制两个轮子的电机本身的驱动特性不可能完全相同，两个电机的外形大小不可能是完全一致，组装时的精度也会出现差异；加之，小车在行驶过程中，轮胎在滚动时可能发生打滑或遇到细小的障碍物等情况，这都会造成左右轮的速度出现差异，从而走不直。本实施例通过实时采集并控制小车左轮电机和右轮电机的PWM值，以此实现小车稳定的直行。
具体地，102、通过预设的控制状态处理器对第一PWM值和第二PWM值进行第一数值调整，得到能够适配预设脉冲神经网络的PWM控制信号。由于小车的PWM值的大小是变化的，为了使PWM值能够适配脉冲神经网络的神经元模型，需要将其值进行缩放。
需要说明的是，第一数值调整采用的缩放方式包括但不限于任何一种已知的数值缩放方式。本实施例以重标度为例。
示例性的，通过控制状态处理器对第一PWM值和第二PWM值进行第一数值调整，将其线性映射到[0,1]的区间，以适配脉冲神经网络的神经元模型。第一数值调整的表达式为：
Figure PCTCN2021137998-appb-000005
其中，x′表示调整后的输出信号，在本实施例中即为PWM调整信号。x表示调整前的输入信号，在本实施例中即为第一PWM值或第二PWM值。min(x)表示输入信号的最小值，在本实施例中即为第一PWM值的最小值或第二PWM值的最小值。max(x)表示输入信号的最大值，在本实施例中即为第一PWM值的最大值或第二PWM值的最大值。
具体地，103、获取预设惯性测量单元记录的小车运动角度的偏移量。在本实施例中，通过惯性测量单元测量小车运动时的偏移量。
惯性测量单元(Inertial measurement unit，简称IMU)是测量物体三轴姿态角(或角速率)以及加速度的装置。陀螺仪及加速度计是IMU的主要元件，其精度直接影响到惯性系统的精度。在实际工作中，由于不可避免的各种干扰因素，而导致陀螺仪及加速度计产生误差，从初始对准开始，其导航误差就随时间而增长，尤其是位置误差，这是惯导系统的主要缺点。所以需要利用外部信息进行辅助，实现组合导航，使其有效地减小误差随时间积累的问题。为了提高可靠性，还可以为每个轴配备更多的传感器。一般而言IMU要安装在被测物体的重心上。惯性测量单元相较于传统的轮速检测装置，对小车型号的适配范围更广，更容易加装到小车上。
具体地，104、通过预设的角度处理器对所述偏移量进行第二数值调整，得到能够适配脉冲神经网络的角度信号。偏移量的处理与PWM值的处理相似，都是为了适应脉冲神经网络的神经元模型。
需要说明的是，第二数值调整采用的缩放方式包括但不限于任何一种已知的数值缩放方式。本实施例以重标度为例。
示例性的，通过角度处理器器对偏移量进行第二数值调整，以适配脉冲神经网络的神经元模型。与第一数值调整不同的是，第二数值调整是将偏移量线性映射到[-1,1]的区间，得到角度信号，以角度信号的正负值来表示小 车的偏移方向。第二数值调整的表达式为：
Figure PCTCN2021137998-appb-000006
其中，θ′表示调整后的输出信号，在本实施例中为角度信号。θ表示调整前的输入信号，在本实施例中为偏移量。max(θ)表示输入信号的最大值，在本实施例即为偏移量的最大值，min(θ)表示输入信号的最小值，在本实施例即为偏移量的最小值，mean(θ)表示偏移量的平均值。
具体地，105、通过脉冲神经网络的有监督学习算法，结合角度信号，对PWM控制信号进行调整，得到左PWM值和右PWM值。该步骤由脉冲神经网络实现，脉冲神经网络以控制状态处理器输出的PWM控制信号为训练信号，以角度处理器输出的角度信号为学习信号，使用相应的脉冲神经网络有监督学习算法作为学习算法进行PWM控制信号的调整。
步骤105如说明书附图4所示，具体包括：
10501、基于预设的神经元模型，以控制状态处理器输出的PWM控制信号为训练信号，结合神经元模型的非线性神经激活函数，得到抽象化的神经元活动；
10502、基于神经元活动和PWM控制信号，得到神经元解码器；
10503、以角度处理器输出的角度信号为学习信号，以脉冲神经网络的有监督学习算法作为学习算法，结合神经元解码器和神经元活动，得到神经元之间的连接权重；
10504、综合神经元活动、神经元解码器和连接权重，得到修正后的左PWM值和右PWM值。
脉冲神经网络(SNN)被认为是第三代神经网络，是更具有生物意义、运行机制最类似大脑的神经网络模型。脉冲神经网络利用不同的神经元模型来进行信息的传递和计算，不同的神经元模型有不同的特性，可以通过更换神经元来寻找更合适的模型。
在步骤10501中，脉冲神经网络中的神经元活动的可抽象表达为：
a＝G[αe·x]
其中，a表示神经元活动，α是与神经元相关的标度因子，e表示神经元的编码器，x表示待编码的信息，即为PWM控制信号。G[·]表示非线性神经激活函数，由具体的神经元模型决定；
在步骤10502中，解码器d可以按照以下公式计算得到：
d＝r -1Τ
T ij＝∫a ia jdx
r＝∫a jxdx
其中，d表示解码器，a i表示神经元i的活动，a j表示神经元j的活动，T和r为中间变量，x表示输入数据，即为PWM控制信号。
在步骤10503中，有监督的学习算法包括多种，可以根据实际需求更换。在本实施例中，有监督学习算法为Prescribed Error Sensitivity有监督学习算法。利用Prescribed Error Sensitivity有监督学习算法得到的权重修正值Δω ij可由以下公式求得：
Figure PCTCN2021137998-appb-000007
其中，Δω ij表示权重修正值，E表示希望最小化的错误向量，在本实施例中即为角度偏移量。κ表示学习率，d表示解码器，n为解码器的维度，ω表示连接权重，α表示与神经元相关的标度因子，e表示神经元的编码器，i、j表示神经元，a i表示神经元i的活动。
综合神经元活动、神经元解码器和连接权重得到修正后的PWM值，具体表达式为：
Figure PCTCN2021137998-appb-000008
其中，a表示神经元活动，ω表示神经元之间的连接权重，d表示神经 元解码器，P表示修正后的PWM值。
将PWM控制信号输入到脉冲神经网络即可得到修正后的PWM值，即左PWM值和右PWM值。脉冲神经网络具有低功耗、动态性好的特点，通过脉冲神经网络结合惯性测量单元(IMU)可以为小车“走不直”的问题提供高度自适应、相对低功耗的解决方案。
具体地，106、将左PWM值和右PWM值分别应用到小车的左轮电机和右轮电机上，调整小车轮速。经过步骤105即可得到左轮和右轮的修正值，将左PWM值和右PWM值应用到小车中，即可调整小车左轮和右轮的移速。
具体地，107、循环上述步骤101-106，实现小车轮速的实时调整。每调整完一次小车的轮速即可采集当前的轮速和偏移量，实现小车轮速的自适应。尤其在复杂环境下，小车时刻受到环境因素的干扰，通过实时调整轮速能够让小车具备更高的适应性。
说明书附图5为本实施例的实验结果图，使用脉冲神经网络和IMU结合后进行小车直走控制的结果。其中，x轴为时间，y轴为角度偏移。通过实验结果可明显看出，经过6s的自适应调整后，小车能够呈直线行走，基本不存在角度偏移。
本实施例提出了一种基于脉冲神经网络的小车轮速自调控方法，利用了脉冲神经网络低功耗、高动态性的特点，将脉冲神经网络和IMU相结合进行小车直走控制，使小车具备轮速自适应能力，解决了小车由于其自身电机特性不相同、外界环境干扰等因素造成的“走不直”问题。相较于现有PID结合轮速检测装置的方案，本实施例提供的方案即使在复杂环境下，也能通过实时调整轮速实现小车直走，具备高度的适应性和抗干扰能力，并且所能适配的智能小车的类型也更为广泛。
实施例2
本实施例提出了一种基于脉冲神经网络的小车轮速自调控系统，采用实施例1提出的一种基于脉冲神经网络的小车轮速自调控方法，小车轮速自调控系统的模块示意图如说明书附图6所示。具体方案如下:
一种基于脉冲神经网络的小车轮速自调控系统，包括循环轮速获取单元1、轮速调整单元2、偏移获取单元3、偏移调整单元4、调整控制单元5、移速应用单元6和实时调整单元7如下：
轮速获取单元1：用于获取小车左轮电机的第一PWM值和小车右轮电机的第二PWM值；
轮速调整单元2：内置有控制状态处理器，用于通过控制状态处理器对第一PWM值和第二PWM值进行第一数值调整，得到能够适配预设脉冲神经网络的PWM控制信号。轮速调整单元2包括对第一PWM值和第二PWM值进行缩放，线性映射到[0，1]区间。
偏移获取单元3：用于获取惯性测量单元记录的小车运动角度的偏移量；
偏移调整单元4：内置有角度处理器，用于通过角度处理器对所述偏移量进行第二数值调整，得到能够适配脉冲神经网络的角度信号。偏移调整单元4包括对偏移量进行缩放，线性映射到[-1，1]区间。
调整控制单元5：内置有脉冲神经网络，用于通过脉冲神经网络的有监督学习算法，结合角度信号，对PWM控制信号进行调整，得到左PWM值和右PWM值；
移速应用单元6：用于将左PWM值和右PWM值分别应用到小车的左轮电机和右轮电机上，调整小车轮速；
实时调整单元7：用于依次循环轮速获取单元1、轮速调整单元2、偏移获取单元3、偏移调整单元4、调整控制单元5和移速应用单元6，实现小车轮速的实时调整。
调整控制单元5包括神经元活动模块51、解码器模块52、连接权重模块53和输出模块54，具体如说明书附图7所示。
神经元活动模块51：用于基于预设的神经元模型，以控制状态处理器输出的PWM控制信号为训练信号，结合神经元模型的非线性神经激活函数，得到抽象化的神经元活动。
脉冲神经网络中的神经元活动的可抽象表达为：
a＝G[αe·x]
其中，a表示神经元活动，α是与神经元相关的标度因子，e表示神经元的编码器，x表示要编码的信息，即为PWM控制信号。G[·]表示非线性神经激活函数，由具体的神经元模型决定。
解码器模块52：用于基于神经元活动和PWM控制信号，得到神经元解码器。
解码器d可以按照以下公式计算得到：
d＝r -1Τ
T ij＝∫a ia jdx
r＝∫a jxdx
其中，d表示解码器，a i表示神经元i的活动，a j表示神经元j的活动，T和r为中间变量，x表示输入数据，即为PWM控制信号。
连接权重模块53：用于以角度处理器输出的角度信号为学习信号，以脉冲神经网络的有监督学习算法作为学习算法，结合神经元解码器和神经元活动，得到神经元之间的连接权重。
在本实施例中，有监督学习算法为Prescribed Error Sensitivity有监督学习算法。利用Prescribed Error Sensitivity有监督学习算法得到的权重修正值Δω ij可由以下公式求得：
Figure PCTCN2021137998-appb-000009
其中，Δω ij表示权重修正值，E表示希望最小化的错误向量，在本实施 例中即为角度偏移量。κ表示学习率，d表示解码器，n为解码器的维度，ω表示连接权重，α表示与神经元相关的标度因子，e表示神经元的编码器，i、j表示神经元，a i表示神经元i的活动。
输出模块54：用于综合神经元活动、神经元解码器和连接权重，得到修正后的左PWM值和右PWM值。
综合神经元活动、神经元解码器和连接权重得到修正后的PWM值，具体表达式为：
Figure PCTCN2021137998-appb-000010
其中，a表示神经元活动，ω表示神经元之间的连接权重，d表示神经元解码器，P表示修正后的PWM值。
将PWM控制信号输入到脉冲神经网络即可得到修正后的PWM值，即左PWM值和右PWM值。脉冲神经网络具有低功耗、动态性好的特点，通过脉冲神经网络结合惯性测量单元(IMU)可以为小车“走不直”的问题提供高度自适应、相对低功耗的解决方案。
本实施例提出了一种基于脉冲神经网络的小车轮速自调控系统，将实施例1的基于脉冲神经网络的小车轮速自调控方法系统化，使其更具实用性。
本发明提供了一种基于脉冲神经网络的小车轮速自调控方法及自调控系统。小车轮速自调控方法利用了脉冲神经网络低功耗、高动态性的特点，将脉冲神经网络和IMU相结合进行小车直走控制，使小车具备轮速自适应能力，解决了小车由于其自身电机特性不相同、外界环境干扰等因素造成的“走不直”问题。相较于现有PID结合轮速检测装置的方案，本发明提供的方案即使在复杂环境下，也能通过实时调整轮速实现小车直走，具备高度的适应性和抗干扰能力，并且所能适配的智能小车的类型也更为广泛。
本领域普通技术人员应该明白，上述的本发明的各模块或各步骤可以 用通用的计算装置来实现，它们可以集中在单个计算装置上，或者分布在多个计算装置所组成的网络上，可选地，他们可以用计算机装置可执行的程序代码来实现，从而可以将它们存储在存储装置中由计算装置来执行，或者将它们分别制作成各个集成电路模块，或者将它们中的多个模块或步骤制作成单个集成电路模块来实现。这样，本发明不限制于任何特定的硬件和软件的结合。
注意，上述仅为本发明的较佳实施例及所运用技术原理。本领域技术人员会理解，本发明不限于这里的特定实施例，对本领域技术人员来说能够进行各种明显的变化、重新调整和替代而不会脱离本发明的保护范围。因此，虽然通过以上实施例对本发明进行了较为详细的说明，但是本发明不仅仅限于以上实施例，在不脱离本发明构思的情况下，还可以包括更多其他等效实施例，而本发明的范围由所附的权利要求范围决定。
以上公开的仅为本发明的几个具体实施场景，但是，本发明并非局限于此，任何本领域的技术人员能思之的变化都应落入本发明的保护范围。

Claims (10)

  1. 一种基于脉冲神经网络的小车轮速自调控方法，其特征在于，包括如下步骤：
    获取小车左轮电机的第一PWM值和右轮电机的第二PWM值；
    通过预设控制状态处理器对所述第一PWM值和所述第二PWM值进行第一数值调整，得到能够适配预设脉冲神经网络的PWM控制信号；
    获取预设惯性测量单元记录的小车运动角度的偏移量；
    通过预设角度处理器对所述偏移量进行第二数值调整，得到能够适配所述脉冲神经网络的角度信号；
    通过所述脉冲神经网络的有监督学习算法，结合所述角度信号，对所述PWM控制信号进行调整，得到左PWM值和右PWM值；
    将所述左PWM值和所述右PWM值分别应用到小车的左轮电机和右轮电机上，调整小车轮速；
    循环上述步骤实现小车轮速的实时调整。
  2. 根据权利要求1所述的小车轮速自调控方法，其特征在于，“通过所述脉冲神经网络的有监督学习算法，结合所述角度信号，对所述轮速值进行调整”具体包括：
    基于预设的神经元模型，以所述控制状态处理器输出的所述PWM控制信号为训练信号，结合所述神经元模型的非线性神经激活函数，得到抽象化的神经元活动；
    基于所述神经元活动和所述PWM控制信号，得到神经元解码器；
    以所述角度处理器输出的角度信号为学习信号，以所述脉冲神经网络的有监督学习算法作为学习算法，结合所述神经元解码器和所述神经元活动，得到神经元之间的连接权重；
    综合所述神经元活动、所述神经元解码器和所述连接权重，得到修正后 的所述左PWM值和所述右PWM值。
  3. 根据权利要求1或2所述的小车轮速自调控方法，其特征在于，所述第一数值调整包括对所述第一PWM值和所述第二PWM值进行缩放，线性映射到[0，1]区间；
    所述第二数值调整包括对所述偏移量进行缩放，线性映射到[-1，1]区间。
  4. 根据权利要求3所述的小车轮速自调控方法，其特征在于，所述第一数值调整的表达式为：
    Figure PCTCN2021137998-appb-100001
    其中，x′表示所述PWM调整信号，x表示所述第一PWM值或所述第二PWM值，min(x)表示所述第一PWM值的最小值或所述第二PWM值的最小值，max(x)表示所述第一PWM值的最大值或所述第二PWM值的最大值；
    所述第二数值调整的表达式为：
    Figure PCTCN2021137998-appb-100002
    其中，θ′表示所述角度信号，θ表示偏移量，max(θ)表示偏移量的最大值，min(θ)表示偏移量的最小值，mean(θ)表示偏移量的平均值。
  5. 根据权利要求2所述的小车轮速自调控方法，其特征在于，所述神经元活动的抽象化表达式为：
    a＝G[αe·x]
    其中，a表示所述神经元活动，α表示与神经元相关的标度因子，e表示神经元的编码器，x表示待编码的信息，G[·]表示由预设神经元模型决定的非线性神经激活函数；
    所述解码器的表达式包括：
    d＝r -1Τ
    T ij＝∫a ia jdx
    r＝∫a jxdx
    其中，d表示解码器，a i表示神经元i的活动，a j表示神经元j的活动，T和r为中间变量，x表示输入数据。
  6. 根据权利要求5所述的小车轮速自调控方法，其特征在于，所述有监督学习算法为Prescribed Error Sensitivity有监督学习算法，得到的权重修正值Δω ij可由以下公式求得：
    Figure PCTCN2021137998-appb-100003
    其中，Δω ij表示权重修正值，E表示希望最小化的错误向量，κ表示学习率，d表示解码器，n为解码器的维度，ω表示连接权重，α表示与神经元相关的标度因子，e表示神经元的编码器，i、j表示神经元，a i表示神经元i的活动。
  7. 根据权利要求6所述的小车轮速自调控方法，其特征在于，综合所述神经元活动、所述神经元解码器和所述连接权重得到修正后的PWM值，具体表达式为：
    Figure PCTCN2021137998-appb-100004
    其中，a表示神经元活动，ω表示神经元之间的连接权重，d表示神经元解码器，P表示修正后的PWM值。
  8. 一种基于脉冲神经网络的小车轮速自调控系统，其特征在于，包括如下：
    轮速获取单元：用于获取小车左轮电机的第一PWM值和小车右轮电机的第二PWM值；
    轮速调整单元：内置有控制状态处理器，用于通过所述控制状态处理器对所述第一PWM值和所述第二PWM值进行第一数值调整，得到能够适配 预设脉冲神经网络的PWM控制信号；
    偏移获取单元：用于获取预设惯性测量单元记录的小车运动角度的偏移量；
    偏移调整单元：内置有角度处理器，用于通过所述角度处理器对进行第二数值调整，得到能够适配所述脉冲神经网络的角度信号；
    调整控制单元：内置有脉冲神经网络，用于通过所述脉冲神经网络的有监督学习算法，结合所述角度信号，对所述PWM控制信号进行调整，得到左PWM值和右PWM值；
    移速应用单元：用于将所述左PWM值和所述右PWM值分别应用到小车的左轮电机和右轮电机上，调整小车轮速；
    实时调整单元：用于依次循环所述轮速获取单元、所述轮速调整单元、所述偏移获取单元、所述偏移调整单元、所述调整控制单元和所述移速应用单元，实现小车轮速的实时调整。
  9. 根据权利要求8所述的小车轮速自调控系统，其特征在于，所述调整控制单元具体包括：
    神经元活动模块：用于基于预设的神经元模型，以所述控制状态处理器输出的所述PWM控制信号为训练信号，结合所述神经元模型的非线性神经激活函数，得到抽象化的神经元活动；
    解码器模块：用于基于所述神经元活动和所述PWM控制信号，得到神经元解码器；
    连接权重模块：用于以所述角度处理器输出的角度信号为学习信号，以所述脉冲神经网络的有监督学习算法作为学习算法，结合所述神经元解码器和所述神经元活动，得到神经元之间的连接权重；
    输出模块：用于综合所述神经元活动、所述神经元解码器和所述连接权重，得到修正后的所述左PWM值和所述右PWM值。
  10. 根据权利要求8或9所述的小车轮速自调控系统，其特征在于，所 述轮速调整单元包括对所述第一PWM值和所述第二PWM值进行缩放，线性映射到[0，1]区间；
    所述偏移调整单元包括对所述偏移量进行缩放，线性映射到[-1，1]区间。
PCT/CN2021/137998 2021-09-10 2021-12-14 基于脉冲神经网络的小车轮速自调控方法及自调控系统 WO2023035462A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111062796.XA CN115782606A (zh) 2021-09-10 2021-09-10 基于脉冲神经网络的小车轮速自调控方法及自调控系统
CN202111062796.X 2021-09-10

Publications (1)

Publication Number Publication Date
WO2023035462A1 true WO2023035462A1 (zh) 2023-03-16

Family

ID=85416788

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/137998 WO2023035462A1 (zh) 2021-09-10 2021-12-14 基于脉冲神经网络的小车轮速自调控方法及自调控系统

Country Status (2)

Country Link
CN (1) CN115782606A (zh)
WO (1) WO2023035462A1 (zh)

Citations (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104950887A (zh) * 2015-06-19 2015-09-30 重庆大学 基于机器人视觉系统和自主跟踪系统的运输装置
CN105305895A (zh) * 2015-11-17 2016-02-03 吉林大学 一种基于转矩反馈和换向补偿的无刷电机控制方法
US20170120755A1 (en) * 2015-10-29 2017-05-04 Kubota Corporation Electric Work Vehicle
CN108427415A (zh) * 2018-04-02 2018-08-21 济南大学 一种现代物流中自动导引避障机器人小车及其控制方法
CN108594829A (zh) * 2018-06-08 2018-09-28 武汉蛋玩科技有限公司 一种智能巡线机器人以及控制方法
CN109085855A (zh) * 2018-09-18 2018-12-25 北京图森未来科技有限公司 一种传感器位置调整装置、系统、调整方法及车辆
CN109240292A (zh) * 2018-09-10 2019-01-18 广州大学 一种基于bp神经网络的自平衡小车及其姿态控制方法
CN110109456A (zh) * 2019-04-25 2019-08-09 杭州师范大学钱江学院 一种小车自动驾驶方法
CN111716359A (zh) * 2020-06-30 2020-09-29 福州大学 复合导航式差速仓储机器人及其工作方法
CN112141210A (zh) * 2020-08-21 2020-12-29 天津市天波科达科技有限公司 一种基于电子差速底盘的航向控制方法及装置
US20210206430A1 (en) * 2020-01-05 2021-07-08 Eva, Llc Automated Steering Control Mechanism and System for Wheeled Vehicles

Patent Citations (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104950887A (zh) * 2015-06-19 2015-09-30 重庆大学 基于机器人视觉系统和自主跟踪系统的运输装置
US20170120755A1 (en) * 2015-10-29 2017-05-04 Kubota Corporation Electric Work Vehicle
CN105305895A (zh) * 2015-11-17 2016-02-03 吉林大学 一种基于转矩反馈和换向补偿的无刷电机控制方法
CN108427415A (zh) * 2018-04-02 2018-08-21 济南大学 一种现代物流中自动导引避障机器人小车及其控制方法
CN108594829A (zh) * 2018-06-08 2018-09-28 武汉蛋玩科技有限公司 一种智能巡线机器人以及控制方法
CN109240292A (zh) * 2018-09-10 2019-01-18 广州大学 一种基于bp神经网络的自平衡小车及其姿态控制方法
CN109085855A (zh) * 2018-09-18 2018-12-25 北京图森未来科技有限公司 一种传感器位置调整装置、系统、调整方法及车辆
CN110109456A (zh) * 2019-04-25 2019-08-09 杭州师范大学钱江学院 一种小车自动驾驶方法
US20210206430A1 (en) * 2020-01-05 2021-07-08 Eva, Llc Automated Steering Control Mechanism and System for Wheeled Vehicles
CN111716359A (zh) * 2020-06-30 2020-09-29 福州大学 复合导航式差速仓储机器人及其工作方法
CN112141210A (zh) * 2020-08-21 2020-12-29 天津市天波科达科技有限公司 一种基于电子差速底盘的航向控制方法及装置

Also Published As

Publication number Publication date
CN115782606A (zh) 2023-03-14

Similar Documents

Publication Publication Date Title
Matveev et al. Nonlinear sliding mode control of an unmanned agricultural tractor in the presence of sliding and control saturation
Ollero et al. Predictive path tracking of mobile robots. Application to the CMU Navlab
CN104942809A (zh) 基于视觉伺服系统的机械臂动态模糊逼近器
US20220306157A1 (en) Vehicle-mounted camera gimbal servo system and control method
CN108919810A (zh) 基于视觉示教的移动机器人定位与导航系统
CN110763224A (zh) 一种自动导引运输车导航方法及导航系统
JP6216545B2 (ja) 自律走行車両、車両の自律走行システム及び自律走行車両の制御方法
CN110109363B (zh) 一种轮式移动机器人编队的神经网络自适应控制方法
WO2023035462A1 (zh) 基于脉冲神经网络的小车轮速自调控方法及自调控系统
CN108594798B (zh) 一种可实现蜂拥控制的机器人小车系统及其控制方法
Raudonis et al. Trajectory based fuzzy controller for indoor navigation
Singh et al. Control of closed-loop differential drive mobile robot using forward and reverse Kinematics
Díaz-Téllez et al. ROS-based Controller for a Two-Wheeled Self-Balancing Robot
D’Ambrosio et al. A new spatial learning control for autonomous vehicles: experimental results
CN113848954A (zh) 一种改进型多神经元神经网络pid算法的无人车伺服控制系统
Liu et al. Goal-driven motion control of snake robots with onboard cameras via policy improvement with path integrals
Gao et al. Path Tracking Control of Micro-tracked mobile robot
CN116811592A (zh) 一种基于脉冲神经网络的小车方向调控方法及其应用
Tak et al. Path Tracing in holonomic drive system with Reduced Overshoot using rotary encoders
CN114312988B (zh) 一种基于偏航角与转向系统转角转换模型的轨迹控制方法
Lin et al. Design of Two-wheeled Self-balancing Traction Vehicle Based on OpenMV
Savnani et al. Modelling, Design and Control of a Four wheel Holonomic Drive
Wang et al. Trajectory tracking of nonholonomic mobile robots via discrete-time sliding mode controller based on uncalibrated visual servoing
CN109131019A (zh) 模式动物自动化运载装置及方法
CN220374669U (zh) 一种行走装置

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: 21956631

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE