Control method, device and storage medium based on parallel learning and extreme value search

Abstract

The present invention relates to a control method, device and storage medium based on parallel learning and extreme value search, wherein the method comprises: step S1: collecting the time series of slip rate and road adhesion coefficient; step S2: constructing the characteristic vector of the braking road surface; step S3: matching the characteristic vector of the braking road surface with the characteristic vector of each known road surface type in the historical database, if there is a known road surface type with a similarity greater than a pre-configured threshold, then executing step S4, otherwise, then executing step S5; step S4: obtaining the optimal slip rate, and outputting a control signal based on an experience control library; step S5: determining the optimal slip rate based on extreme value search, and inputting it into an adaptive control model based on parallel learning to obtain a control signal; step S6: controlling based on the obtained control signal. Compared with the prior art, the present invention has the advantages of having the shortest braking distance when emergency braking across road types.

Description

Technical Field

[0001] The present invention relates to the field of intelligent driving, and in particular to a control method, device and storage medium based on parallel learning and extreme value search. Background Art

[0002] Nonlinear systems are widely present in all walks of life. For example, as vehicles develop towards intelligence and unmanned driving in the future, research on intelligent emergency braking control that can autonomously identify road types and achieve the best emergency braking effect is a prerequisite for the practical implementation of driverless cars and is also the key to the safety of people's lives and property. Since the friction and braking torque of tire-road contact are nonlinear, and due to the diversity of road conditions, the friction and braking torque of different road types in contact with tires present different nonlinear types, making it a huge challenge for driverless cars to achieve the best emergency braking effect under various road conditions, especially variable road conditions.

[0003] The extreme value search and extreme value following control of nonlinear systems such as emergency braking have always been an important direction in the control field. In particular, when the type of nonlinear system exists, changes or switches, its extreme value will also change. How to ensure the stability of the system when the nonlinear system type changes or switches and make the system run quickly and stably at the new extreme value is the difficulty of nonlinear type change and switching nonlinear system control. Summary of the invention

[0004] The purpose of the present invention is to provide a control method, device and storage medium based on parallel learning and extreme value search.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A control method based on parallel learning and extreme value search, comprising:

[0007] Step S1: When the vehicle is in an emergency braking state, the time series of the slip rate and the road adhesion coefficient are collected;

[0008] Step S2: constructing a characteristic vector of the braking road surface based on the collected time series of the slip rate and the road adhesion coefficient;

[0009] Step S3: Based on the feature vector of the braking road surface, match the feature vectors of each known road surface type in the historical database. If there is a known road surface type with a similarity greater than a pre-configured threshold, execute step S4; otherwise, execute step S5;

[0010] Step S4: obtaining an optimal slip ratio corresponding to the known road surface type, and outputting a control signal based on an empirical control library;

[0011] Step S5: determining the optimal slip ratio based on extreme value search, and inputting the optimal slip ratio into an adaptive control model based on parallel learning to obtain a control signal output by the adaptive control model;

[0012] Step S6: Perform control based on the obtained control signal.

[0013] The feature vector is a binary data sequence, and each element in the feature vector consists of a slip rate and a corresponding road adhesion coefficient.

[0014] The determining of the optimal slip ratio based on extreme value search comprises:

[0015] Extreme value search step: determine the optimal slip rate of the current road surface based on the extreme value search algorithm;

[0016] When the optimal slip ratio is obtained, the switch is set to the off state, the extreme value search is stopped and the optimal slip ratio is output;

[0017] When the switch is in the off state, the differential of the road adhesion coefficient and the differential of the slip rate difference are calculated. If the absolute value of the differential of the difference is greater than a preconfigured first threshold, the switch is set to the on state and the extreme value search step is returned.

[0018] The differential of the road adhesion coefficient and the differential of the slip rate difference are obtained by a tracking differentiator.

[0019] The tracking-differentiator uses the fastest comprehensive function to obtain the differential of the adhesion coefficient and the differential of the slip rate difference.

[0020] The control signal u output by the adaptive control model is:

[0021]

[0022] Where: G λ (v) is the control gain of the controlled object, is the derivative of the optimal slip rate, K is the control gain, sgn() is the sign function, e is the difference between the optimal slip rate and the current slip rate, g is the acceleration of gravity, v is the vehicle speed, λ is the slip rate, J is the moment of inertia of the tire, R is the tire radius, m is one-quarter of the vehicle mass, is the neural network coefficient, Φ(λ) is the radial basis function.

[0023] The radial basis function is a Gaussian function.

[0024] The control signal u is output based on the experience control library i for:

[0025]

[0026] in: is the derivative of the optimal slip rate for a known road type i, is the optimal slip rate of the known road type i, is the neural network coefficient of the known road type i, and k is the number of known road types.

[0027] A control device based on parallel learning and extreme value search comprises a memory, a processor, and a program stored in the memory. When the processor executes the program, the method as described above is implemented.

[0028] A storage medium stores a program, which implements the above method when executed.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] 1. The optimal slip rate and the corresponding control signal are quickly obtained based on the characteristic vector matching of the road surface, which has the advantage of fast braking force response. In addition, for unknown roads, the braking force response rate of unknown roads can be improved based on extreme value search and parallel learning.

[0031] 2. An extreme value search mechanism with a switching switch is adopted, which can avoid the jump of the output optimal slip rate when the road surface type remains unchanged, and can respond and output the latest optimal slip rate in time when the road surface type changes.

[0032] 3. The fastest comprehensive function is used to obtain the differential of the adhesion coefficient and the differential of the slip rate difference, which can avoid the problem of too large differential signal error caused by the large noise caused by the derivation of the actual signal.

[0033] 4. A neural network based on radial basis function is used to approximate the nonlinear curve of the tire-road model. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a schematic flow chart of the main steps of the method of the present invention;

[0035] Figure 2 It is a technical logic block diagram of the present invention;

[0036] Figure 3 Schematic diagram of the improved extreme value search mechanism of the present invention. DETAILED DESCRIPTION

[0037] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0038] A control method based on parallel learning and extreme value search, such as Figure 1 and Figure 2 As shown, including:

[0039] Step S1: When the vehicle is in an emergency braking state, the time series of the slip rate and the road adhesion coefficient are collected, wherein the feature vector is a binary data sequence, and each element in the feature vector consists of the slip rate and the corresponding road adhesion coefficient.

[0040] Step S2: constructing a characteristic vector of the braking road surface based on the collected time series of the slip rate and the road adhesion coefficient;

[0041] Step S3: Based on the feature vector of the braking road surface, match the feature vectors of each known road surface type in the historical database. If there is a known road surface type with a similarity greater than a pre-configured threshold, execute step S4; otherwise, execute step S5;

[0042] Step S4: obtaining an optimal slip ratio corresponding to the known road surface type, and outputting a control signal based on an empirical control library;

[0043] Step S5: determining the optimal slip ratio based on extreme value search, and inputting the optimal slip ratio into an adaptive control model based on parallel learning to obtain a control signal output by the adaptive control model;

[0044] In this embodiment, an improved extreme value search mechanism is adopted. Specifically, determining the optimal slip rate based on the extreme value search includes:

[0045] Extreme value search step: determine the current optimal slip rate based on the extreme value search algorithm;

[0046] When the optimal slip ratio is obtained, the switch is set to the off state, the extreme value search is stopped and the optimal slip ratio is output;

[0047] When the switch is in the off state, the differential of the road adhesion coefficient and the differential of the slip rate difference are calculated. If the absolute value of the differential of the difference is greater than a preconfigured first threshold, the switch is set to the on state and the extreme value search step is returned.

[0048] like Figure 3As shown, compared with the traditional extreme value search mechanism, the newly added parts of this embodiment mainly include a switching judgment condition and a tracking-differentiator. The switching judgment condition is mainly used to avoid jitter after searching for the extreme value and to determine whether the nonlinear type has changed. Since the road adhesion coefficient differential and the slip rate difference and their differentials are required in the switching judgment condition, directly deriving the signal to obtain the differential signal in actual applications will cause large errors due to the presence of noise, etc. To solve this problem, this embodiment draws on the method of extracting differential signals by the tracking-differentiator to realize the differential acquisition of the road adhesion coefficient differential and the slip rate difference. The switching judgment condition designed in this embodiment is as follows:

[0049] 1.

[0050] 2. The duration is less than 0.1s.

[0051] in, is the differential of the road adhesion coefficient and the differential of the slip rate difference, ρ is the excitation signal, which is a triangular wave or a sine signal. When the above two conditions are met at the same time, it means that the extreme value of the corresponding road adhesion coefficient has been searched.

[0052] The differential of the road adhesion coefficient and the differential of the slip ratio difference are obtained by a tracking-differentiator.

[0053] The tracking-differentiator uses the fastest synthesis function to obtain the differential of the adhesion coefficient and the differential of the slip rate difference.

[0054] The specific calculation is as follows:

[0055] (1)

[0056] The specific calculation formula of formula (1) is as follows:

[0057] (2)

[0058] in , For status, , is the function control parameter, is the simulation or sampling step size. Quickly catch up with input signal without overshoot , for The approximate differential of , the differential signal of the tracking process. Different from the object sampling period of Function step size. Calculated by formula (3):

[0059] fhan( x 1 ( k )- e ( k ), x 2 ( k ), r 0 , h 0 ) = - r [ a d - sign ( a )] s a - r 0 sign ( a ) (3)

[0060] The corresponding parameters in formula (3) can be calculated by formula (4).

[0061] d = r 0 h 0 a 0 = h 0 x 2 y = x 1 + a 0 a 1 = d ( d +8 y ) a 2 = a 0 + sign ( y ) ( a 1 - d ) 2 s y = [ sign ( y + d )- sign ( y - d )] 2 a =( a 0 + y - a 2 ) s y + a 2 s a = [ sign ( a + d )- sign ( a - d )] 2 (4)

[0062] Obtained by calculation It can be approximated as .

[0063] Parallel learning can meet the requirement of continuous excitation by using the recorded historical data in parallel with the current data. In this embodiment, parallel learning forms a closed loop with the controlled object, and the learned nonlinear type model and its extreme values ​​are recorded in the experience control library and the historical database. In the follow-up control, the input-output mapping curve of the nonlinear system model is approximated by a neural network based on radial basis function:

[0064] (5)

[0065] in: is the adhesion coefficient, is the slip rate, The slip rate is the input of the RBF neural network, and the vector weight is W = [ ω 1 , ω 2 ,..., ω p ] p ∈ R p , Represents the number of neuron nodes of the RBF neural network, The value can be set based on experience. Φ ( λ ) = [ ϕ 1 ( λ- ξ 1 ), ϕ 2 ( λ- ξ 2 ),..., ϕ p ( λ- ξ p )] T is the regression vector composed of radial basis functions, is the nonlinear mapping function radial basis function of the hidden layer, is the neuron (node) position of the RBF neural network. The Gaussian function is selected as the radial basis function of the hidden layer:

[0066] (6)

[0067] Since neural networks have a wide range of approximation capabilities, as long as enough RBF neural network nodes are arranged, any continuous function can be approximated. To accurately model, the RBF neural network needs to meet the continuous excitation conditions:

[0068] (7)

[0069] in , is a constant value, The advantage of parallel learning is that it can accurately model the unknown model while ensuring the adaptive follow-up of the target value. Based on the requirements of this application for emergency braking control, the adaptive control design of parallel learning is the mode of formula (8):

[0070] (8)

[0071] Where: G λ (v) is the control gain of the controlled object, is the derivative of the optimal slip rate, K is the control gain, sgn() is the sign function, e is the difference between the optimal slip rate and the current slip rate, g is the acceleration of gravity, v is the vehicle speed, λ is the slip rate, J is the moment of inertia of the tire, R is the tire radius, m is one-quarter of the vehicle mass, is the neural network coefficient, Φ(λ) is the radial basis function.

[0072] Used to approximate the adhesion coefficient-slip rate curve. The weight update is as follows:

[0073] (9)

[0074] in: is the control input; Control gain for the controlled object; is the vehicle speed; is the moment of inertia of the tire; One quarter of the vehicle mass; is the tire radius; is the optimal slip ratio; To control the gain; is to adjust the gain; the error signal is ; Weight update formula (9) This is the focus of parallel learning. It can be observed that the continuous motivation condition can be met by selectively recording data to establish a historical database. The specific data selection algorithm is shown in Table 1:

[0075] Table 1 Data selection algorithm

[0076] if #timg# then#timg##timg# end ifif #timg# then#timg##timg# for #timg# do #timg##timg##timg# end forif #timg# then#timg##timg# else#timg# #timg# end ifend if

[0077] As shown in Table 1, the upper part is mainly to selectively record data to meet the continuous excitation conditions when the recorded data does not meet the continuous excitation conditions. is the radial basis function transformation of the last selected data, is the set error value, Indicates that when the difference or distance between the latest data and the last data is greater than a certain value, the current data will be recorded in the historical database When the recorded data is sufficient to meet the continuous excitation condition, the data selection algorithm enters the second stage. The main function of the second stage is to speed up the approximation of the model by selecting data that can maximize the singular value. After the modeling is completed, the nonlinear type experienced this time is recorded in the form of fixed weights of the neural network. Recorded in the experience control library, and the corresponding extreme values ​​are recorded in it for future retrieval.

[0078] When faced with experienced road types, the optimal slip rate following control for known road types mainly includes three parts, in addition to the controlled object, the decision mechanism, the experience control library and the decision mechanism. The model of the experience control library is mainly obtained by the adaptive following control module based on parallel learning. The experience control library stores the road model and the optimal slip rate value of the corresponding road type, which are:

[0079] W ̂ iT Φ ( λ ) = [ W ̂ 1 iT Φ ( λ ),..., W ̂ p iT Φ ( λ )] T , （ i =1 ,..., k ） (10)

[0080] λ d i = [ λ d 1 ,..., λ d k ] T , （ i =1 ,..., k ） (11)

[0081] In the formula Representative Known road surface types, The number of known road surface types stored. The value can increase as the number of road surface types increases. Representative The optimal slip ratio value of the road surface type is obtained. The empirical control library is established based on the above equations (10) and (11), as shown in equation 12:

[0082] (12)

[0083] in: is the derivative of the optimal slip rate for a known road type i, is the optimal slip rate of the known road type i, is the neural network coefficient of the known road type i.

[0084] Different from formula (8), formula (12) directly calls the corresponding neural network weights and optimal slip ratio values ​​instead of calculating them, so the control efficiency is high and the response speed is fast.

[0085] The decision-making mechanism is mainly based on equations (13) and (14):

[0086] (13)

[0087] (14)

[0088] if and Is the setting value, representing the road surface type and model When the road surface type is close enough, it is considered that the current road surface type is The road surface of the model is the same, and the optimal slip ratio value and optimal slip ratio following controller of the corresponding model are called. and When , it indicates that the current road type is new, and the control is switched to the optimal slip rate following control for unknown road types based on parallel learning.

[0089] A control device based on parallel learning and extreme value search comprises a memory, a processor, and a program stored in the memory. When the processor executes the program, the above method is implemented.

[0090] A storage medium stores a program, which implements the above method when executed.

[0091] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.

Claims

1. A control method based on parallel learning and extreme value search, characterized in that: include: Step S1: When the vehicle is in an emergency braking state, the time series of the slip rate and the road adhesion coefficient are collected; Step S2: constructing a characteristic vector of the braking road surface based on the collected time series of the slip rate and the road adhesion coefficient; Step S3: Based on the feature vector of the braking road surface, match the feature vectors of each known road surface type in the historical database. If there is a known road surface type with a similarity greater than a pre-configured threshold, execute step S4; otherwise, execute step S5; Step S4: obtaining an optimal slip ratio corresponding to the known road surface type, and outputting a control signal based on an empirical control library; Step S5: determining the optimal slip ratio based on extreme value search, and inputting the optimal slip ratio into an adaptive control model based on parallel learning to obtain a control signal output by the adaptive control model; Step S6: performing control based on the obtained control signal; The determining of the optimal slip ratio based on extreme value search comprises: Extreme value search step: determine the optimal slip rate of the current road surface based on the extreme value search algorithm; When the optimal slip ratio is obtained, the switch is set to the off state, the extreme value search is stopped and the optimal slip ratio is output; When the switch is in the off state, the differential of the road adhesion coefficient and the differential of the slip rate difference are calculated. If the absolute value of the differential of the difference is greater than a preconfigured first threshold, the switch is set to the on state and the extreme value search step is returned.

2. A control method based on parallel learning and extreme value search according to claim 1, characterized in that: The feature vector is a binary data sequence, and each element in the feature vector consists of a slip rate and a corresponding road adhesion coefficient.

3. A control method based on parallel learning and extreme value search according to claim 1, characterized in that: The differential of the road adhesion coefficient and the differential of the slip rate difference are obtained by a tracking differentiator.

4. A control method based on parallel learning and extreme value search according to claim 3, characterized in that: The tracking-differentiator uses the fastest comprehensive function to obtain the differential of the adhesion coefficient and the differential of the slip rate difference.

5. The control method based on parallel learning and extreme value search according to claim 1, characterized in that: The control signal u output by the adaptive control model is: ， Where: G λ (v) is the control gain of the controlled object, is the derivative of the optimal slip rate, K is the control gain, sgn() is the sign function, e is the difference between the optimal slip rate and the current slip rate, g is the acceleration of gravity, v is the vehicle speed, λ is the slip rate, J is the moment of inertia of the tire, R is the tire radius, m is one-quarter of the vehicle mass, is the neural network coefficient, Φ(λ) is the radial basis function.

6. A control method based on parallel learning and extreme value search according to claim 5, characterized in that: The radial basis function is a Gaussian function.

7. A control method based on parallel learning and extreme value search according to claim 5, characterized in that: The control signal u is output based on the experience control library i for: , in: is the derivative of the optimal slip rate for a known road type i, is the optimal slip rate of the known road type i, is the neural network coefficient of the known road type i.

8. A control device based on parallel learning and extreme value search, comprising a memory, a processor, and a program stored in the memory, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 7 is implemented.

9. A storage medium having a program stored thereon, characterized in that: When the program is executed, the method according to any one of claims 1 to 7 is implemented.

Images