JPH10285707A - Drive control apparatus for electric vehicle - Google Patents

Abstract

PROBLEM TO BE SOLVED: To obtain a drive control apparatus, coping with the cause of a control error and which enhances the reliability and the accuracy of a control operation by installing a during-a-run learning management means which gives a teacher signal to a neural network for update, during the running operation of a vehicle on the basis of an adjustable-speed request, on the basis of an input signal which indicates a vehicle-body motion state and on the basis of the output of a neural network for control. SOLUTION: Torque commands TR, TL which are generated in a neural network 312 for control are output to a motor control part. On the other hand, they are input to a learning control part 604 installed inside a learning management part 600. The learning control part 604 generates torque commands TRS, TLS as teacher signals on the basis of the torque commands TR, TL and on the basis of input signals γt, β, δt, T so as to be given to a neural network 602 for update. By the teacher signals, a learning operation is performed regarding the generation of the torque command TR, TL.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a drive control device mounted on an electric vehicle provided with motors corresponding to respective drive wheels and controlling each motor so as to increase the running stability of the electric vehicle.

[0002]

2. Description of the Related Art For conventional engine vehicles, so-called four-wheel steering (4WS) has been developed in order to achieve and improve the performance of smoothly performing acceleration / deceleration, turning, lane change and the like, that is, running stability. I have. However, the current 4
In WS, a mechanism related to steering and an electronic control system (or a hydraulic control system in some cases) for controlling the mechanism are slightly large-scale and complicated, so that a large mounting space is required, the response is slightly slow, There are problems such as failures easily occurring. On the other hand, running stability is also required for an electric vehicle that is a vehicle propelled by a vehicle-mounted motor. It is conceivable to mount a 4WS for a conventional engine vehicle on an electric vehicle, but in such a case, the above-described problems cannot be avoided. Also, especially with pure electric vehicles that only have a motor as the propulsion source of the vehicle, it is common to install a much larger battery than that of a conventional engine vehicle, so the mounting space problem is It becomes even more noticeable.

Under such circumstances, various studies have been made on running stability control means or methods that can be mounted on an electric vehicle. For example, in JP-A-1-298903, JP-A-1-298904, JP-A-1-298905 and JP-A-5-176418, a plurality of motors are arranged so that each of the left and right driving wheels can be rotated independently. With respect to an electric vehicle of the type, that is, an electric vehicle with left and right driving wheels independently driven, a method of performing motor torque control or speed control in accordance with a difference in travel locus between left and right driving wheels is disclosed. In the left and right driving wheel independent driving type electric vehicle, since different torques can be generated in the left and right driving wheels,
Similar effects to 4WS in conventional gasoline vehicles
Without large and complicated mechanisms and control systems such as Note that a typical example of a type capable of performing the control described in each of the above publications is an electric vehicle in which a vehicle driving motor is embedded in or integrated with each of left and right drive wheels, that is, a wheel-in motor type electric vehicle. . The wheel-in motor type electric vehicle has a motor for each driving wheel, so there is no need for a distribution mechanism such as a differential gear, so low transmission loss and low energy consumption can be used. There are advantages such that the motor can be made smaller as compared with a so-called one-motor type electric vehicle as long as it can be driven, and that since the motor is built in the drive wheel, the integration is high and the vehicle interior space is widened.

[0004]

However, only the difference between the running trajectories of the left and right driving wheels means the rotational movement of the vehicle body around the center of gravity.
It is not possible to sufficiently cope with the movement of the vehicle body in the lateral direction, slippage of the drive wheels, etc., and it is not possible to achieve sufficient running stability. For example, when steering is performed for turning, various amounts of yaw rate, lateral acceleration, slip angle, and the like are generated in the vehicle body. These are quantities whose values are determined according to the vehicle speed, the wheel speed, and the like in addition to the cornering force, the lateral force, the road surface friction coefficient, and the like, and the values cannot be uniquely estimated from the traveling locus difference.

In order to solve such deficiencies, the present applicant has already proposed a technique for controlling the output (torque) of each motor so that the response of the yaw rate or the slip angle to the steering angle matches the target response ( Japanese Patent Application No. 9-8693
issue). By using such a control method, that is, target yaw rate adaptation control or slip angle adaptation control (for example, zero slip angle control), it is possible to appropriately cope with rotational motion, lateral motion, and the like, thereby improving the running stability of the vehicle. Further, the steering angle is an amount that is immediately generated with the steering by the vehicle operator, and in the prior proposal, since the control based on this steering angle is performed, the effect of reducing the frequency of the pedal operation by the vehicle operator also occurs. That is, even if the vehicle operator steers while driving with the accelerator pedal, and as a result the traveling becomes unstable, before the vehicle operator senses it and starts to depress the brake pedal, Since the target yaw rate adaptation control and the target slip angle adaptation control are activated and the running instability is eliminated or reduced, the operation of the brake pedal by the vehicle operator becomes unnecessary, and the running with the accelerator pedal depressed becomes possible.

The inventor has been studying to further improve the drive control device according to the prior proposal. One of the points that need to be improved as a result of this study is that if control is performed based solely on the steering angle, the steering angle will be too large or too small, disturbances after steering (road surface, cross wind, etc.), delays in the control system, etc. Due to
An unexpected disturbance (control error) occurs in the vehicle body motion. Another point that needs to be improved is to model the vehicle body motion in the form of a basic body motion equation and perform target yaw rate adaptation control or target slip angle adaptation control according to the control logic derived based on this model. The control error due to the difference between the vehicle and this model cannot be eliminated, and various fluctuation factors (e.g., changes in vehicle characteristics over time, changes in loaded weight, etc.) in an actual use environment cannot be dealt with.

A first object of the present invention is to provide a control for maintaining and improving the running stability of a vehicle, that is, a running stability control in a left-right driving wheel independent driving type electric vehicle represented by a wheel-in motor type electric vehicle. By introducing a neural network, it is possible to deal with various control error factors, including control error factors that cannot be expressed only by the basic body motion equation of the vehicle body and therefore cannot be reflected in the control logic. The purpose is to improve performance and accuracy. A second object of the present invention is to further improve the reliability, accuracy, responsiveness, and the like of the running stability control by performing parallel and distributed processing of the running stability control using the neural network and updating the organization of the neural network. It is in. A third object of the present invention is to update the organization of a neural network for driving stability control based on the results of learning from acceleration / deceleration requests and vehicle body movement conditions actually generated during driving, thereby improving the reliability of control. And to further improve the accuracy.

[0008]

According to a preferred embodiment of the present invention, a plurality of motors for individually driving the driving wheels of an electric vehicle are individually given commands relating to their outputs. This is an embodiment of a drive control device that controls the running of an electric vehicle, and relates to generation of the above-mentioned command according to an acceleration / deceleration request and a vehicle body motion state, and a control that has been learned before the start of vehicle running based on an actual vehicle test or a simulation calculation result. A neural network for updating and a neural network for updating, and a learning learning management means for providing a teacher signal to the neural network for updating while the vehicle is traveling based on an input signal indicating an acceleration / deceleration request and a vehicle body motion state and an output of the neural network for control. It is characterized by having. The control neural network generates and outputs a command to be individually given to each of the plurality of motors according to the input signal while the vehicle is running. The neural network for updating, based on the input signal indicating the acceleration / deceleration request and the vehicle body movement state and the output of the control neural network, i.e., the instruction to the motor, during the running of the vehicle, receives the command corresponding to the acceleration / deceleration request and the vehicle body movement state. Learn about generating. The updating neural network updates the organization of the controlling neural network based on the result of learning. This tissue update transmits the synapse coupling coefficient and the neuron threshold value in the updating neural network to the control neural network, and thereby at least partially updates / replaces the synaptic coupling coefficient and the neuron threshold value in the control neural network. For example, it can be executed even while the vehicle is running.

As described above, in the present embodiment, a command to the motor is issued based on the contents learned in advance by the control neural network based on the results of an actual vehicle test or a simulation operation instead of the basic body motion equation of the vehicle body in which the vehicle body motion is modeled. Since it is generated, there is no control error due to the difference that would occur between the model and the actual vehicle motion when the vehicle motion is modeled. Also, since the control neural network can be updated while the vehicle is running,
It is possible to quickly and accurately cope with various factors such as an excessively large or small steering angle, a disturbance received after steering, a delay in a control system, and the like in an actual use environment. Therefore, control reliability, accuracy, responsiveness, and the like are improved.

The present invention can be expressed not only as a "drive control device" but also as a "drive control method", "neutral network use method", "electric vehicle", and the like. Any such changes to the wording will be subject to the parties referring to this disclosure.
It can be done easily.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Preferred embodiments of the present invention will be described below with reference to the drawings.

(1) System Configuration FIG. 1 shows a system configuration of an electric vehicle suitable for carrying out the present invention. This electric vehicle is a rear wheel drive, and a right rear wheel 10RR and a left rear wheel 10RL are drive wheels, and a right front wheel 10RR.
FR and the left front wheel 10FL are driven wheels. However, the present invention can be applied to a front wheel drive vehicle and a four wheel drive vehicle.

The electric vehicle shown in FIG. 1 is a wheel-in motor type. That is, the vehicle driving motor 12R on the right side in the drawing
Is the left vehicle driving motor 1 inside the right rear wheel 10RR.
2L is incorporated inside the left rear wheel 10RL. FIG. 2 shows an example of a mounting method using the right rear wheel 10RR as an example. In this figure, the rotor 18 is fixed inside the wheel 16 so as to be able to rotate integrally with the tire 14, while the stator 24 is fixed to the rear axle 22 via the motor shaft 20, and the stator 24 is connected to the rotor 18 via bearings or the like. Has been joined. Further, the stator 2 is provided on the inner wall surface of the rotor 18.
While a rotor magnet (permanent magnet) 26 is fixed so as to oppose the motor 4 with a small gap, a stator winding 28 is wound around the stator 24, and a cable 30 for flowing a current through the stator winding 28 is connected to the motor shaft. 20 to the stator winding 28. In such a structure, by supplying an alternating current to the stator winding 28 via the cable 30, the rotor 18 rotates and generates propulsion of the vehicle. In practicing the present invention, other types of structures may be used, or other types of structures provided with motors corresponding to the left and right driving wheels may be used.

The battery 32 shown in FIG.
It is a drive power supply source to the 2R and 12L, and its discharge output is supplied to the motor 12R via the inverter 34R and to the motor 12L via the inverter 34L. Inverters 34R and 34L are a type of power converter. That is, the inverter 34R is connected to the motor control unit 3
Under the control of 6R, the discharge output (DC) of the battery 32 is converted into a power format (three-phase AC in this figure) suitable for the motor 12R, and the inverter 34L controls the discharge output of the battery 32 under the control of the motor control unit 36L. Is converted into a power format suitable for the motor 12L. The motor control unit 36R controls the inverter 34R in response to a torque command TR from the vehicle control unit 38.
Is controlled to output a torque corresponding to the torque command TR from the motor 12R. Similarly, the motor control unit 36L controls the inverter 34L according to the torque command TL from the vehicle control unit 38 to output a torque corresponding to the torque command TL from the motor 12L. The motor control units 36R and 36L also have a function of insulating and separating the corresponding one of the inverters 34R and 34L from the vehicle control unit 38. Further, the inverter 3 is controlled by the motor control units 36R and 36L.
The control of 4R and 34L is based on each phase current detection value of the motors 12R and 12L obtained from a current sensor (not shown),
Alternatively, the motors 12R and 1R determined from the rotor angular position and the like
This is performed based on the estimated value of each phase current of 2L. In addition, the present invention
It can be applied not only to a pure electric vehicle but also to a so-called hybrid vehicle.

The vehicle controller 38 includes motors 12R and 12R.
The control member is responsible for controlling the output torque of L, monitoring and controlling the status of each component on the vehicle, notifying the vehicle occupants of the vehicle status, and other functions. It can be realized by hardware-like modification. Outputs of sensors provided in various parts of the vehicle are input to the vehicle control unit 38, and the vehicle control unit 38 uses the outputs of the sensors for motor output control and vehicle state monitoring.

For example, a wheel speed sensor 40RR (for example, a resolver) provided on the wheel (16 in FIG. 2) of the right rear wheel 10RR outputs a signal (for example, every minute angular displacement) indicating the wheel speed VRR of the right rear wheel 10RR. The wheel speed sensor 40RL provided on the wheel of the rear left wheel 10RL generates a signal indicating the wheel speed VRL of the rear left wheel 10RL. Similarly, the wheel speed sensor 40FR provided on the wheel of the right front wheel 10FR is
A wheel speed sensor 40FL provided on the wheel of the front left wheel 10FL generates a signal indicating the wheel speed VFR of the left front wheel 10FL. The accelerator sensor 42 outputs a signal indicating the amount of depression of an accelerator pedal (not shown), that is, the accelerator opening VA, and the brake sensor 44 transmits a signal indicating the amount of depression of the brake pedal 56, that is, a braking force FB, to a shift position switch. 46 is
A signal indicating a closing range of a shift lever (not shown) (and a shift lever position within the range in an engine brake range or the like), that is, a shift position is generated. Further, the steering angle sensor 48 outputs a signal indicating a steering angle δt that changes according to the operation of a steering wheel (not shown), the yaw rate sensor 50 outputs a signal indicating a yaw rate γt acting on the vehicle body, and the lateral acceleration sensor 54 outputs a signal indicating the yaw rate γt acting on the vehicle body. Are generated, respectively, indicating the lateral acceleration Gy acting on.
The outputs of these sensors are each converted into data in a format that can be processed by the vehicle control unit 38 when input to the vehicle control unit 38. The vehicle control unit 38 uses the converted data to determine the torque commands TR and TL, switch control methods, and the like.

In FIG. 1, a braking system for hydraulically braking the front wheels and regeneratively braking the rear wheels is used. That is,
When the brake pedal 56 is depressed, the hydraulic pressure generated in the master cylinder 58 is transmitted to the left and right wheel cylinders 60R and 60L in response to the depression, and the left and right brake wheels 6
2R and 62L, and while the braking torque is applied to the left and right front wheels, the vehicle control unit 38 controls the regeneration torque according to the braking force (for example, the hydraulic pressure of the master cylinder 58) FB detected using the brake sensor 44. Commands TR and T
L is generated. Therefore, as shown in FIG. 3, the distribution of the braking force in the vehicle of FIG. 1 is a distribution in which both the hydraulic regeneration increases as the braking force FB (“pedal input” on the horizontal axis) increases. As described above, since the hydraulic system and the regenerative system are separated from each other after the brake sensor 44, even if one of the hydraulic pressure and the regeneration fails, the vehicle can be evacuated by the other. In addition, since the hydraulic system is not provided with mechanisms such as valves and pumps and the electric system for driving and controlling the same, the system configuration is simpler than, for example, a system that shuts off the hydraulic pressure during recuperation. Become.
It should be noted that mechanisms such as valves and pumps and their driving and
One of the reasons why there is no need to provide an electric system for control is that the running stability control is performed using the control of the output torque of the motors 12R and 12L as described later. is there.

(2) Overall Function of Vehicle Control Unit FIG. 4 shows a functional configuration of the vehicle control unit 38 in the present embodiment. Although the present invention can be implemented by hardware logic or software, the following description uses a block diagram for clarifying a functional configuration. The vehicle control unit 38 includes a torque command provisional determination unit 100, a TRC / ABS equivalent control unit 200, a traveling stability control unit 300, a control switching unit 400, a switching control unit 500, and a learning management unit 600. Torque command provisional determination unit 1
00 captures the outputs of the accelerator sensor 42, the brake sensor 44, the shift position switch 46, and the like as information indicating a request from the vehicle operator, and outputs information on the rotation speeds NR and NL of the motors 12R and 12L to be controlled to the wheels. The torque required to generate an output corresponding to a request such as acceleration or deceleration from the vehicle operator under the current motor rotation speed is taken in from the speed sensors 40RR and 40RL, and the left and right driving wheels 10RR are generated based on the information. And 10RL, and the obtained torques are used as torque commands TR and TL. However, at this stage, the torque commands TR and TL are not finally determined, but are only "tentatively determined". Torque command provisional determination unit 10
0 is a torque command T further determined based on the vehicle speed VS.
(Described later) is also generated.

The TRC / ABS equivalent control section 200 corrects the provisionally determined torque commands TR and TL under specific conditions, thereby obtaining the TRC / AB of a conventional engine vehicle.
An operation having an effect equivalent to S is executed. The running stability control unit 300 controls the torque commands TR and T based on the torque command T from the torque command provisional determination unit 100 and the output of each sensor.
L is generated, and the torque command generates a moment M around the center of gravity of the vehicle body to prevent the running from becoming unstable during steering. The learning management unit 600 updates the organization of the neutral network in the running stability control unit 300 during running. Both the TRC / ABS equivalent control unit 200 and the traveling stability control unit 300 utilize the characteristic of the left and right wheel independent drive type electric vehicle that the torque commands can be individually given to the left and right motors 12R and 12L, respectively. This operation is performed. Therefore, in the present embodiment, it is possible to provide a function that contributes to the running stability of the vehicle, including a function corresponding to TRC / ABS in a conventional engine vehicle, without a hydraulic control system or a complicated electronic control system.

Control switching section 400 and switching control section 500
Selectively switches the control logic according to the vehicle state or the driving state. The control switching section 400 is provided after the TRC / ABS equivalent control section 200 and the traveling stability control section 300, and the switching control section 500 gives a control signal to the control switching section 400. That is, whether to output the torque command TR and TL provisionally determined by the torque command provisional determination unit 100 and corrected by the TRC / ABS equivalent control unit 200, or whether the torque command TR and TL generated by the traveling stability control unit 300 Whether to output TL is determined by the control switching unit 400 based on a control signal from the switching control unit 500.

(3) Function of torque command provisional determination section As shown in FIG. 5, the torque command provisional determination section 100 determines whether or not the accelerator pedal is depressed based on the output of the accelerator sensor 42. It has a part 102. When the on / off determining unit 102 determines that the accelerator is on, that is, the accelerator pedal is being depressed, the accelerator opening calculating unit 104 calculates the accelerator opening VA based on the output of the accelerator sensor 42, and the accelerator is off, that is, the accelerator is off. When it is determined that the pedal has not been depressed, the braking force calculation unit 106 calculates a braking force (stepping force) FB based on the output of the brake sensor 44. At this time, the shift position set by the shift position switch 46 is referred to by the accelerator opening calculator 104 and the braking force calculator 106, and the accelerator opening VA and the braking force FB are calculated and determined according to the shift position at that time. Is done. The calculated and determined accelerator opening VA is supplied to powering torque calculators 108, 110 and 112, and the braking force FB is supplied to regenerative torque calculators 114, 116 and 118.

On the other hand, the wheel speed calculator 122 converts a signal, for example, a pulse signal from the wheel speed sensor 40RR into a wheel speed VRR.
The wheel speed calculation unit 124 converts the signal from the wheel speed sensor 40RL into data indicating the wheel speed VRL, and the wheel speed calculation unit 126 converts the signal from the wheel speed sensor 40F.
The signal from R is converted into data indicating the wheel speed VFR, and the wheel speed calculation unit 128 converts the signal from the wheel speed sensor 40FL into data indicating the wheel speed VFL. The rotation speed calculation unit 130 arranged downstream of the wheel speed calculation unit 122
By multiplying by 0 / (2πR), the wheel speed VRR is converted into the rotation speed NR, and the rotation speed calculation unit 132 arranged at the subsequent stage of the wheel speed calculation unit 124 is multiplied by 60 / (2πR) to obtain the wheel speed VRR. The VRL is converted into the rotation speed NL. The vehicle speed calculating unit 134 disposed after the wheel speed calculating units 126 and 128 is a vehicle speed calculating unit.

VS = (VFR + VFL) / 2 is obtained, and the rotation speed calculation unit 136 arranged at the subsequent stage of the vehicle speed calculation unit 134 multiplies the vehicle speed VS by 60 / (2πR) to convert the vehicle speed VS to the rotation speed N. Convert. Here, the rotation speed is expressed in rpm and the wheel speed is expressed in m / sec. R is a wheel radius.

Of the rotational speed information thus obtained, the rotational speed NR of the right motor 12R is calculated by the power running torque calculator 108 and the regenerative torque calculator 114, and the rotational speed NL of the left motor 12L is calculated by the power running torque calculation. The unit 110, the regenerative torque calculator 116, and the average rotational speed N corresponding to the vehicle speed VS are input to the powering torque calculator 112 and the regenerative torque calculator 118, respectively.

The right running torque calculation unit 108 refers to the running torque map 138 with the rotation speed NR of the right motor 12R as a key, thereby obtaining the rotation speed NR at that time.
, The maximum powering torque that can be output from the right motor 12R is determined, and the maximum powering torque is divided by the accelerator opening VA to determine the torque command TR for the right motor 12R. As shown in FIG. 6, the powering torque map 138 used here is a means for holding the rotation speed-maximum torque characteristic in a powering region (a region where the rotation speed> 0 and the torque> 0). By referring to the rotational speed NR as a key, the rotational speed N
The maximum powering torque at R (the point on the curve of VA = 100% in FIG. 6) is obtained. Further, assuming that the accelerator opening VA at that time is x%, the power running torque to be output, that is, the accelerator opening VA and the rotation at that time, are calculated by proportional processing of multiplying the obtained maximum power running torque by x / 100. A torque command TR (point on the curve of VA = x% in FIG. 6) corresponding to the number NR can be obtained. The powering torque calculation unit 110 on the left side also obtains a powering torque to be output, that is, a torque command TL corresponding to the accelerator opening VA and the rotational speed NL at that time in the same procedure. Powering torque calculation unit 112
In the same manner, the torque command T corresponding to the accelerator opening VA and the rotation speed N at that time, that is, information obtained by converting the vehicle speed VS into a torque command is generated.

The right regenerative torque calculation unit 114 refers to the regenerative torque map 140 with the right motor 12R rotation speed NR as a key, thereby obtaining the rotation speed NR at that time.
, The maximum regenerative torque that can be output from the right motor 12R is obtained, and the maximum regenerative torque is divided by the braking force FB to determine the torque command TR for the right motor 12R. As shown in FIG. 7, the regenerative torque map 140 used here has a regenerative region (rotation speed>
0 and torque <0), which is a means for retaining the rotational speed versus maximum torque characteristic in the region where the rotational speed NR is used as a key as described above.
(The point on the curve of FB = 100% in FIG. 7). Further, if the braking force FB at that time is x%, the regenerative torque to be output, that is, the braking force FB and the rotation speed NR at that time, are proportionally processed by multiplying the obtained maximum regenerative torque by x / 100.
(The point on the curve of FB = x% in FIG. 7) can be obtained. The regenerative torque calculator 116 on the left side also obtains a regenerative torque to be output, that is, a torque command TL corresponding to the braking force FB and the rotational speed NL at that time, in a similar procedure. The regenerative torque calculation unit 118 also generates information in which the torque command T, that is, the vehicle speed VS is converted into a torque command based on the braking force FB and the rotation speed at that time.

The powering / regeneration switching unit 120 switches between commanding the powering torque and the commanding the regenerative torque according to the output of the on / off determining unit 102. That is,
Whether the torque commands TR, TL, and T determined by the powering torque calculators 108, 110, and 112 are supplied to the subsequent stage, or whether the regenerative torque calculators 114, 116, and 11 are provided.
Whether the torque commands TR, TL, and T determined in step 8 are supplied to the subsequent stage is switched according to the on / off state of the accelerator. Hereinafter, for convenience of description, the torque command provisional determination unit 1
The torque commands TR and TL determined at 00 are referred to as “temporarily determined” torque commands TR and TL. T is not a command to the motor itself, but is referred to as a torque command for convenience.

In FIG. 5, the left and right power running torque maps 1
38 and the regenerative torque map 140 are shared, but this is an example in the case where the left and right motors 12R and 12L have the same characteristics, and when they do not have the same characteristics, separate maps are used for the left and right. 6 and 7, only the maximum power running torque or the maximum regenerative torque is mapped.
A large number of torque curves may be mapped using the accelerator opening VA or the braking force FB as a parameter. If such a mapping is performed, the torque command can be obtained by referring to the map based on the pair of the rotation speed and the accelerator opening or the braking force, so that the powering or the regenerative torque calculation units 108 to 118 There is no need to perform pro-rata processing. On the other hand, mapping only the number of rotations versus the maximum torque characteristic can save storage space for holding the map. Furthermore, it is simpler to execute the calculation of the torque commands TR and TL in a time-sharing manner by a single arithmetic unit, instead of simultaneously and concurrently using the left and right individual functions and arithmetic members. In order to adopt such a configuration, for example, a timing clock that alternately provides the execution timing of the calculation for the left driving wheel and the execution timing of the calculation for the right driving wheel is generated, and the calculation is performed in synchronization with the timing clock. do it. The same applies to the calculation of various coefficients and correction amounts described later.

(4) Function of Switching Control Unit The control switching unit 400 shown in FIG. 4 specifically includes the torque commands TR and T generated by the running stability control unit 300.
L is output to the motor control units 36R and 36L, or the TRC /
Torque command T corrected by ABS equivalent control unit 200
Switching between outputting R and TL is performed. That is,
The provisionally determined torque commands TR and TL are corrected by the TRC / ABS equivalent control unit 200 only when the wheel angular acceleration satisfies a predetermined condition, and output to the motor control units 36R and 36L. This switching of the control logic is executed according to a control switching signal supplied from the switching control unit 500.

[0029]

[Table 1] Table 1 shows the control switching signal generation logic employed in the switching control section 500. As shown in this table, the switching control unit 500 determines that at least one of the angular acceleration dωR / dt of the right rear wheel 10RR and the angular acceleration dωL / dt of the left rear wheel 10RL becomes equal to or greater than the predetermined threshold value TH. Assuming that a slip or a tendency has occurred in one of the left and right drive wheels, the torque commands TR and TL corrected by the TRC / ABS equivalent control unit 200 are applied to the motor control units 36R and 36R.
A control switching signal is generated so as to be output to 6L. When the angular accelerations dωR / dt and dωL / dt are both below the threshold value TH, the switching control unit 500
A control switching signal is generated so that the torque commands TR and TL generated by the traveling stability control unit 300 are output to the motor control units 36R and 36L. In the present embodiment, therefore, a function corresponding to TRC or ABS in a conventional engine vehicle can be realized by electrical information processing from detection of wheel speeds VRR and VRL to output of torque commands to motor control units 36R and 36L. No mechanism for controlling the oil pressure is required.

In order to realize the above control switching logic, for example, means for detecting and determining the angular accelerations dωR / dt and dωL / dt, and based on the result of this determination and Table 1
Any means may be provided for generating a control switching signal in accordance with the above logic. An example of the switching control unit 500 provided with these means,
As shown in FIG. In this figure, the wheel acceleration calculation unit 506 determines the acceleration dVRR / dt of the right rear wheel 10RR by differentiating the wheel speed VRR of the right rear wheel 10RR detected by the wheel speed calculation unit 122, and calculates the angular acceleration calculation unit 510. Calculates the angular acceleration dωR / dt of the right rear wheel 10RR by dividing the acceleration dVRR / dt by the wheel radius R, and the slip determination unit 514 compares the angular acceleration dωR / dt with the threshold TH. Similarly, the wheel acceleration calculation unit 508 obtains the acceleration dVRL / dt of the left rear wheel 10RL by differentiating the wheel speed VRL detected by the wheel speed calculation unit 124.
The angular acceleration calculation unit 512 divides the acceleration dVRL / dt by the wheel radius R to obtain the angular acceleration dω of the left rear wheel 10RL.
L / dt, and the slip determination unit 516 determines the angular acceleration dω
L / dt is compared with a threshold value TH. The control operation selecting section 518 generates a control switching signal based on the determination results in the slip determining sections 514 and 516 and according to the logic shown in Table 1. With such a configuration, control switching according to the above-described logic is realized.

(5) Function of TRC / ABS equivalent control unit

[Table 2] The TRC / ABS equivalent control unit 200 calculates the angular acceleration dωR /
Based on dt and dωL / dt, the feedback torque ΔTR for the right rear wheel 10RR and the feedback torque ΔTL for the left rear wheel 10RL are obtained according to the logic of Table 2, and the feedback torque Δ is calculated from the temporarily determined torque command TR.
TR and the feedback torque ΔTL is subtracted from the temporarily determined torque command TL. The threshold values THω1 to THω3 appearing in the left half of Table 2 and the gains G1 and G2 and the correction terms S1 and S2 appearing in the right half depend on the state of the motors 12R and 12L and according to the characteristics.
The C / ABS equivalent control unit 200 sets this. Threshold value TH
ω1 to THω3, gains G1 and G2, and correction term S1
And S2 are made dependent on the state and characteristics of the motors 12R and 12L.
For example, even in a vehicle for driving in an urban area where the output torques of the 2R and 12L change over a wide range, the TRC / ABS equivalent control can be started at a suitable timing and the TRC / A
A torque suitable for ABS equivalent control can be generated. That is, JP-A-2-299402 and JP-A-3-299402
As compared with the forklift described in Japanese Patent No. 27701, an electric vehicle capable of suitably suppressing or eliminating slip can be obtained. For the function, operation, and effect of the TRC / ABS equivalent control unit 200, refer to Japanese Patent Application Laid-Open No. 8-182119 and Japanese Patent Application No. 9-8693, which are assigned to the same applicant as the present application.

FIG. 9 shows a TRC / ABS equivalent control unit 200.
1 shows an example configuration. In the figure, the right angular acceleration determination unit 202
Sets the angular acceleration dωR / dt to a threshold value THω1 to THω
3 (where THω1>THω2>0> THω3), the angular acceleration dωR / dt exceeds THω1,
It is determined which region belongs to THω1 or less and THω2 or more, THω2 or less and THω3 or less, and THω3 or less. The feedback torque calculator 204 calculates the angular acceleration d
When ωR / dt belongs to the region exceeding THω1, the feedback torque calculation unit 206 outputs the angular acceleration dωR / d
When t is less than THω1 and greater than THω2, the feedback torque calculation unit 208 outputs the angular acceleration dω
When R / dt belongs to the region of THω2 or less and exceeds THω3, the feedback torque calculation unit 210 outputs the feedback torque for the right rear wheel 10RR when the angular acceleration dωR / dt belongs to the region of THω3 or less. Calculate ΔTR. Feedback torque calculation unit 2
The arithmetic expressions used in 04 to 210 are shown in Table 2 below.
Are the first, second, third, and fourth equations from the top of the right half of FIG. The adder 212 generates a corrected torque command TR by subtracting the feedback torque ΔTR from the temporarily determined torque command TR, and outputs the corrected torque command TR to the motor control unit 36R.

Similarly, in the figure, the left angular acceleration determination unit 214
Sets the angular acceleration dωL / dt to a threshold value THω1 to THω
3, the angular acceleration dωL / dt becomes T
Over Hω1, THω1 or below THω2, THω2 or below TH
It is determined to which of the regions exceeding ω3 and THω3 or less belongs. The feedback torque calculator 216 determines that the feedback torque is higher when the angular acceleration dωL / dt belongs to a region where THω1 or more and THω2 or less and THω2 or less. Arithmetic unit 220 determines that the angular acceleration dωL / dt belongs to the region of THω2 or less and exceeds THω3, and feedback torque arithmetic unit 222 determines that the angular acceleration dωL / dt belongs to the region of THω3 or less. The feedback torque ΔTL related to the wheel 10RL is calculated. The arithmetic expressions used in the feedback torque arithmetic units 216 to 222 are the fifth, sixth, seventh, and eighth expressions from the top of the right half of Table 2, respectively (signs are in the same order). The adder 224 calculates the feedback torque ΔT from the provisionally determined torque command TL.
By reducing L, a corrected torque command TL is generated and output to the motor control unit 36L.

Threshold values THω1 to THω3, gain G1
And G2 and the correction terms S1 and S2 are determined by the rotational speed determining units 226 and 228 and the coefficient determining unit 230.
232 and threshold value calculation units 234 and 236. In this case, the coefficient determination map 238 is used. Among these, the rotation speed determination unit 226 compares the rotation speed NR of the motor 12R calculated by the rotation speed calculation unit 130 with the base rotation speed NB. Base rotation speed N here
B is a constant torque region where the torque of the motor 12R (and 12L) is restricted by the constant torque line, as shown in FIGS. 6, 7 and 10, and the torque of the motor 12R (and 12L) is the constant power line. Is the rotation speed indicating the boundary with the constant power region limited by the following. The coefficient determination unit 230 turns on / off information on whether the accelerator is on or off.
In addition to the input from the off determination unit 102, information on whether or not the rotation speed NR exceeds the base rotation speed NB is input from the rotation speed determination unit 226, and based on these inputs,
It is determined which of the regions 1 to 4 in FIG. 10 the current operating point of the motor 12R belongs to. Area 1 is a constant torque power running area, 2 is a constant power power running area, 3 is a constant torque regeneration area, and 4 is a constant power regeneration area. Similarly, the rotation speed determination unit 228 compares the rotation speed NL of the motor 12L calculated by the rotation speed calculation unit 132 with the base rotation speed NB. The coefficient determination unit 232 inputs information on whether the accelerator is on or off from the on / off determination unit 102, and determines the information on whether the rotation speed NL exceeds the base rotation speed NB by the rotation speed determination. Unit 228, and based on these inputs, the current motor 12L
It is determined to which of the areas 1 to 4 the operating point belongs.

The coefficient determining section 230 obtains a coefficient value to be used in the area to which the operating point of the motor 12R currently belongs from the coefficient determining map 238. Coefficient determination unit 232
Acquires the coefficient value to be used in the area to which the operating point of the motor 12L currently belongs from the coefficient determination map 238. The coefficient determination map 238 stores various coefficients in association with the area to which the motor operating point belongs. More specifically, on the coefficient determination map 238, as shown in FIG. 10, the coefficients a1 to a3, b1 to b3, c1 to c3, and d1 for determining the threshold values THω1 to THω3.
To d3 and coefficients A1, A2, B1, B2, C1, C2 for determining the feedback gains G1 and G2.
D1 and D2 are generally set to different values for each area. Further, the correction terms S1 and S2 are generally set to different values depending on whether the vehicle is running or regenerating. Right threshold calculator 234
Is calculated using the determined coefficient

(When the accelerator is on) THω1 = a1 × exp (b1 × VA + c1 × VRR +
d1 x VA x VRR) TH? 2 = a2 x exp (b2 x VA + c2 x VRR +
d2 × VA × VRR THω3 = a3 × exp (b3 × VA + c3 × VRR +
d3 × VA × VRR) G1 = A1 × exp (B1 × VA + C1 × VRR +
D1 x VA x VRR) G2 = A2 x exp (B2 x VA + C2 x VRR +
D2 × VA × VRR (when the accelerator is off) THω1 = a1 × exp (b1 × FB + c1 × VRR +
d1 × FB × VRR THω2 = a2 × exp (b2 × FB + c2 × VRR +
d2 × FB × VRR THω3 = a3 × exp (b3 × FB + c3 × VRR +
d3 × FB × VRR G1 = A1 × exp (B1 × FB + C1 × VRR +
D1 × FB × VRR G2 = A2 × exp (B2 × FB + C2 × VRR +
D2 × FB × VRR), the threshold values THω1 to THω1
ω3 and feedback gains G1 and G2 are determined. Similarly, the left threshold value calculation unit 236 calculates the threshold value THω in a formula using VRL instead of VRR in the following formula.
1 to THω3 and feedback gains G1 and G2 are determined. Note that VRR is obtained from the wheel speed calculation unit 122 by
L is input from the wheel speed calculator 124, VA is input from the accelerator opening calculator 104, and FB is input from the brake force calculator 106.

As described above, the feedback torques ΔTR and ΔTL are determined by using the threshold value, the feedback gain, and the correction term determined according to the characteristics of the motors 12R and 12L.
Therefore, even in an application in which the characteristics of the motors 12R and 12L are used over a wide range, it is necessary to generate appropriate feedback torques ΔTR and ΔTL according to the states of the motors 12R and 12L and the requirements of the vehicle operator. Can be. Further, the feedback torques ΔTR and ΔT
The torque commands TR and TL corrected using L are output to the motor control units 36R and 36L when at least one of the angular accelerations dωR / dt and dωL / dt is greater than or equal to the threshold value TH. , Slip or the tendency thereof can be suitably prevented or suppressed. In this embodiment, since it is assumed that the left and right motors 12R and 12L have the same characteristics, the left and right coefficient determination maps 23
8 are shared, but when the characteristics are not the same, separate maps may be used.

In this embodiment, the angular acceleration dωR /
Although dt and dωL / dt are determined as thresholds, the threshold may be determined as slip ratio = | (wheel speed of drive wheels−vehicle speed) / (wheel speed of drive wheels) |. However, TRC / A
In the embodiment for executing the switch to the BS equivalent control,
Since the angular acceleration of the driving wheel is used in the TRC / ABS equivalent control, the angular acceleration of the driving wheel is more preferable in terms of reducing the calculation load. The wheel speed can be obtained by detecting the wheel speeds of the left and right driven wheels and calculating the average value.

(6) Functions of the Running Stability Control Unit and the Learning Management Unit FIG. 11 shows the running stability control unit 300 and the learning management unit 600.
The functional configuration of is shown. The running stability control unit 300 controls the torque command T obtained by the powering torque calculating unit 112 or the regenerative torque calculating unit 118 so that running instability does not occur on the vehicle body during running of the vehicle or running instability is reduced.
Has the function of generating the torque commands TR and TL to be supplied to the motor control units 36R and 36L based on.

In order to realize such a function, the running stability control unit 300 receives the torque command T as described above, while outputting the yaw rate sensor 50, the lateral acceleration sensor 54, the steering angle sensor 48, and calculating the vehicle speed. The vehicle speed VS obtained by the section 134 is input. The running stability control unit 300
A yaw rate calculation unit 302 that converts an output signal of the yaw rate sensor 50 into data indicating the yaw rate γt, a lateral acceleration calculation unit 304 that converts an output signal of the lateral acceleration sensor 54 into data indicating the lateral acceleration Gy, and an output of the steering angle sensor 48 Steering angle calculator 3 for converting a signal into data indicating a steering angle δt
06. Further, the lateral acceleration Gy, the yaw rate γ
Based on t and the vehicle speed VS,

## EQU3 ## A slip angular velocity calculator 308 for calculating a slip angular velocity dβ / dt based on dβ / dt = Gy / VS-γt, and a slip angular velocity dβ / dt obtained by the slip angular velocity calculator 308 are integrated with time. Is provided with a slip angle calculation unit 310 that obtains a slip angle β by the following formula. The various information thus obtained, namely, the yaw rate γt and the slip angle β indicating the motion state of the vehicle body, the steering angle γt indicating the turning request from the vehicle operator, and the torque command indicating the acceleration / deceleration request from the vehicle operator T is
The signals are input to the control neural network 312 provided in the running stability control unit 300 and the updating neural network 602 provided in the learning management unit 600, respectively. The control neural network 312 and the update neural network 602 include:
In any case, the learning of the generation of the torque commands TR and TL according to γt, β, δt, and T is performed before the vehicle starts running, and the input signals γt, β, δt are input during the vehicle running.
And T, the output signals TR and TL are generated.

The torque commands TR and TL generated by the control neural network 312 are supplied to the motor control units 36R and 36L in principle. However, in order to prevent the torque commands TR and TL generated by the control neural network 312 from being out of the output possible range of the motors 12R and 12L, a torque limit calculation unit 31 is provided downstream of the control neural network 312.
4 is provided, and the torque command supply from the running stability control unit 300 to the motor control units 36R and 36L is restricted by the torque restriction calculation unit 314. The torque limit calculation unit 314 can be realized using, for example, a map having the contents shown in FIG. Needless to say, the output of the TRC / ABS equivalent control unit 200 may be used instead of the torque commands TR and TL output from the traveling stability control unit 300 as described above.

While the torque commands TR and TL generated by the control neural network 312 are output to the motor control units 36R and 36L in this way, the learning control unit provided in the learning management unit 600 604 is also input. The learning control unit 604 generates torque commands TRS and TLS as teacher signals based on the torque commands TR and TL and the input signals γt, β, δt, and T to each neural network, as described later, Give to network 602. The updating neural network 602 learns the generation of the torque commands TR and TL according to the input signals γt, β, δt, and T while the vehicle is running, based on the teacher signal. The updating neural network 602 further controls the synaptic coupling coefficients Wji and Vkj and the neuron thresholds θj and λk obtained as a result of the learning by the controlling neural network 3.
12 to update the synaptic coupling coefficient and the neuron threshold value in the control neural network 312.

FIG. 13 conceptually shows a functional configuration of the control neural network 312 and the update neural network 602. As shown in this figure, the neural network for control 312 and the neural network for updating 602 each have an input layer 312a or 602.
a, an intermediate layer 312b or 602b and an output layer 312c or 602c. In addition, in this figure, i
-Th (i = 1 to 4) input layer neuron 312d or 6
02d to j-th (j = 1 to 5) hidden neurons 3
Synaptic connection 312g or 6 leading to 12e or 602e
The coefficient of 02g is represented as Wji, and the output neuron 312f or 60k of the kth (k = 1 or 2) from the jth intermediate layer neuron 312e or 602e.
The coefficient of the synaptic connection 312h or 602h leading to 2f is represented as Vkj.

In the present embodiment, since there are four types of input signals and two types of output signals, the input layer 312a or 60
Neuron 312d or 602d provided in 2a
Is four, and the number of neurons 312f or 602f provided in the output layer 312c or 602c is two. However, when implementing the present invention, it is also possible to omit one of the four types of input signals used in the present embodiment or to add another type of input signal. In the present embodiment, the intermediate layer 312b
Alternatively, the number of neurons 312e or 602e in the intermediate layer 312b or 602b is five, and the number of neurons 312e or 602e in the intermediate layer 312b or 602b is special. No limitation is required. Furthermore, in this figure, a neuron whose own output is one of the inputs is not used, but such a neuron may be used in practicing the present invention.

FIG. 14 shows a learning procedure executed by the control neural network 312 and the updating neural network 602 before the vehicle starts running (before shipping, etc.) under the control of a device (not shown). As shown in this figure, at the time of learning, first, the synaptic coupling coefficient W
ji and Vkj, hidden neurons 312e or 60
The threshold value θj of 2e and the threshold value ηk of the output layer neuron 312f or 602f are initially set to random numbers (700). However, the random numbers used for this initialization are generated within a relatively small range.

After the initialization, the control neural network 312 and the update neural network 6
02 inputs a learning pattern from an external device (not shown) (702). Here, the learning pattern is obtained by associating a combination of the input signals γt, β, δt, and T with the torque commands TR and TL to be obtained from the output layer 312c or 602c according to the input. The control neural network 312 and the update neural network 602 include a synapse coupling coefficient and a neuron threshold value initialized in step 700 and an input signal γ constituting a learning pattern input in step 702.
Based on t, β, δt and T,

## EQU4 ## The output Hj of the hidden neurons 312e and 602e is calculated according to the following equation: Uj = ΣWjiｊIi + θi Hj = f (Uj) (704). Where Σ represents the sum of i, and f is the sigmoid function

F (x) = 1 / {1 + exp (-x)}. The control neural network 312 and the update neural network 602 then follow the following equation:

The output Ok of the output layer neuron 312f or 602f is calculated according to the following equation: Sk = ＳVkj · Hj + ηk Ok = f (Sk) (706). Where 中
Is the sum for j.

Thus, the output layer neuron 312f
Or after calculating the output Ok of 602f, the control neural network 312 and the update neural network 602 determine the torque commands TR and TL constituting the learning pattern input in step 702 as the teacher signal Tk,

The output error λk of the output layer neuron 312f or 602f is calculated according to λk = (Ok−Tk) · Ok · (1−Ok) (708). Control neural network 312 or update neural network 60
2, then the following equation

The output error σj of the hidden neuron 312e or 602e is calculated according to σj = Σλk · Vkj · Hj · (1-Hj) (710). Where 中
Is the sum for k.

The control neural network 312 and the update neural network 602 use the output errors λk and σj obtained in this way to generate synaptic coupling coefficients Wji and Vkj and a neuron threshold θj.
And ηk are updated. That is, the following equation

Vkj = Vkj + ε · λk · Hj ηk = ηk + φ · λk The synaptic coupling coefficient Vkj and the neuron threshold λk are updated according to the following equation (712).

## EQU10 ## The synaptic coupling coefficient Wji and the neuron threshold value .theta.j are updated in accordance with Wji = Wji + .epsilon..sigma.j.Ii.theta.j = .theta.j + .phi..sigma.j (714). Here, ε and φ are constants.

After updating the synaptic coupling coefficients Wji and Vkj and the neuron threshold values θj and ηk in this manner, the control neural network 312 and the update neural network 602 differ from the learning pattern input in step 702. A learning pattern is selected (716), and the operation after step 702 is executed again. As a result of repeatedly executing the operations of steps 702 to 714 while updating the learning pattern in this manner,
When the learning for all the learning patterns prepared in advance is completed (718), the neural network for control 312 and the neural network for updating 602 are reached.
Stores the completion of one type of learning, and executes learning for all learning patterns again (720).
When the number of repetitions of learning for all learning patterns reaches the predetermined limit number (722), the control neural network 312 and the update neural network 602 end the learning procedure to be executed before the vehicle starts running.

The control neural network 31
2 and the update neural network 602, the learning procedure described above may be executed. However, only one of the control neural network 312 and the update neural network 602 executes the learning, and the process proceeds to steps 700 to 722. When the learning is completed, the result may be transferred to the remaining one neural network (724). In this way, it is possible to eliminate the difference between the two neural networks at the start of vehicle running (at the time of shipment) with respect to the synaptic coupling coefficients Wkj and Vkj and the neuron thresholds θj and ηk at the start of running of the vehicle.

The learning method in steps 700 to 722 in FIG. 14 is called an error back propagation method (back propagation method). That is, an output signal obtained by giving an input signal of a certain combination to the neural network is compared with an output signal that is assumed in advance, that is, a teacher signal, and a difference between the output signal and the teacher signal is obtained.

Equation 11 is a E = Σ (Tk-Ok) 2/2 a continue to update the synaptic weight coefficients and neuron threshold to minimize techniques. In the above equation, Σ represents the sum of k. To minimize the error E of the output signal Ok with respect to the teacher signal Tk with respect to the synaptic coupling coefficient Vkj,

Vkj may be updated so that dE / dVkj = dE / dOk · dOk / dSk · dSk / dVkj = (Ok−Tk) · Ok · (1−Ok) · Hj becomes smaller. In the above equation 9, the coefficient of Hj appearing on the right side of equation 12 is set as λk. Similarly, to minimize the error E with respect to the synaptic coupling coefficient Wji,

DE / Wji = dE / dSk · dSk / dHj · dHj / dUj · dUj / dWji = −ｄλk · Vkj · Hj · (1−Hj) · Ii In the above equation, Σ represents the sum regarding k. Equation 8 above is obtained by setting the coefficient of Ii appearing on the right side of Equation 13 as σj.

FIGS. 15 to 17 show a control neural network 312 and an update neural network 602.
4 shows a flow of an operation of the learning control unit 604 during traveling of the vehicle. These procedures are repeatedly executed by the corresponding members at predetermined intervals. One of the features of this embodiment is that the control neural network 312 executes the torque commands TR and TL while the vehicle is running, based on the result of learning performed before the vehicle starts running, as described below. , Update neural network 602
Performs learning based on the actual input signal under the control of the learning control unit 604, and executes the control neural network 312.
Is updated sequentially with the tissue of the updating neural network 602 obtained by learning, such as the synaptic connection coefficient or the neuron threshold value.

First, as shown in FIG. 15, the control neural network 312 inputs a yaw rate γt, a slip angle β, a steering angle δt, and a torque command T (80).
0), the output Hj of the hidden neuron 312e and the output layer neuron 3 based on the same principle as in steps 704 and 706.
The output Ok of 12f is calculated in order (802, 804).
The control neural network 312 outputs the output Ok obtained by the calculation, that is, the torque commands TR and TL, to the motor control units 36R and 36R via the torque limit calculation unit 314.
While supplying to 6L, it supplies to learning control part 604.
As shown in FIG. 17, the learning control unit 604 inputs the yaw rate γt, the slip angle β, the steering angle δt, and the torque command T in addition to the torque commands TR and TL generated by the control neural network 312 ( 1000),
The teacher signal TRS related to the torque command TR and the teacher signal TLS related to the torque command TL are expressed by the following equations.

TRS = FR (γt, β, δt, T) TLS = FL (γt, β, δt, T) is generated (1004), and the teacher signals TRS and TL are generated.
S is supplied to the updating neural network 602 (1006).

The updating neural network 602 is
In addition to the yaw rate γt, the slip angle β, the steering angle δt, and the torque command T, the teacher signals TRS and TLS are input (FIG. 16: 900). Update Neural Network 6
02, the output Hj of the intermediate layer neuron 602e, the output Ok of the output layer neuron 602f, and the output layer neuron 602f by the same procedure as steps 704 to 710.
And the output error σj of the hidden neuron 602e are calculated in order (902-908). The updating neural network 602 determines in steps 712 and 71
The synapse coupling coefficients Wji and Vkj and the neuron thresholds θj and λk are updated by the same procedure as in step 4 (910, 912). The updating neural network 602 calculates the updated synaptic coupling coefficients Wji and Vkj.
Then, the neuron threshold values θj and λk are transferred to the control neural network 312 (914). The control neural network 312 calculates the synaptic coupling coefficient W
ji and Vkj and neuron thresholds θj and λk
Is transmitted from the updating neural network 602 (806), the own synaptic coupling coefficients Wji and Vkj and the neuron thresholds θj and λk are updated based on the transferred information (80).
8).

Therefore, in the present embodiment, the torque commands TR and TL are generated using the control neural network 312 based on the input signals yaw rate γt, slip angle β, steering angle δt and torque command T, and the wheel angular acceleration Are not large (that is, as long as slip or the tendency is not considered to occur), the torque commands TR and T
The motors 12R and 12L are controlled using L.
The organization of the control neural network 312 is as follows:
It is formed before the vehicle starts running by the procedure shown in FIG. Therefore, when executing the procedure shown in FIG. 14, if the data obtained by running the actual vehicle or the data obtained by the numerical simulation (simulation calculation) is set in the learning pattern, only the basic motion equation of the vehicle body is sufficient. In addition, it is possible to cope with factors that cannot be taken into the control model, and therefore, the reliability and accuracy of control are improved as compared with the related art. Further, while the vehicle is traveling, the torque commands TR and TL generated by the control neural network 312 and the input signals γt, φ, δt, and T under the actual traveling environment are set.
The learning control unit 604 uses the learning signals TRS and TL
S is generated, and the update neural network 602 transfers the result of learning performed based on the teacher signals TRS and TLS to the control neural network 312 to update the organization. The organization of the control neural network 312 can be updated according to the acceleration / deceleration request and the vehicle body motion state,
More reliable control will be realized. Furthermore,
Since the torque command generation processing by the control neural network 312 and the learning by the update neural network 602 are performed simultaneously and in parallel, the speed of control is increased due to the efficiency of the processing.

Further, in this embodiment, the amount used as an input signal to each neural network is:
In any case, the amount is an amount appearing in the basic motion equation of the vehicle body or an amount that can be derived from the amount. First, considering that there is no left-right difference between the wheels and the right-left difference in the cornering force, the basic motion equation of the vehicle body is

(15) m · V · (dφ / dt + γt) = 2 · (YF + YR) ··· The lateral movement of the vehicle body, where YF = YFR + YFL, YR = YRR + YRL I · dγt / dt = 2 · (LF · YF-LR · YR) + M: Movement in the yaw direction around the center of gravity of the vehicle body. See FIG. 18 for the variable names and their meanings. On the other hand, assuming that the equation of motion of the rotation of each wheel has no difference between the left and right wheels and that the commanded torque is exactly the actual output as it is,

(Equation 16) Becomes See FIG. 19 for the variable names and their meanings. In addition, the cornering force

YF = −KF · {φ + (LF / VS) · γt−δt} YR = −KR · {φ + (LR / VS) · γt} Therefore, by substituting Equations 16 and 17 into Equation 15, and rearranging,

S · φ (s) = A11 · φ (s) + A12 ·
γT (s) + B11 · δt (s) + C11 · M (s) s · γt (s) = A21 · φ (s) + A22 · γt
(S) + B21 · δt (s) + C21 · M (s) where A11 = −2 · (KF + KR) / (m · VS) A12 = −1-2 · (LF · KF−LR · KR) / (m
^{· VS 2) B11 = 2 ·} KF / (m · VS) C11 = 0 A21 = -2 · (LF · KF-LR · KR) / I A22 = -2 · (LF 2 · KF + LR 2 · KR) / ( I
VS) The simultaneous differential equation of B21 = 2 LF KF / I C21 = 1 / I is obtained. Where s is the Laplace operator.

Among the quantities used as input signals in the present embodiment, the yaw rate γt, the slip angle β and the steering angle δt appear in the simultaneous differential equations shown in Expression 18. The torque command T is expressed in the same formula in the form of the vehicle speed VS and the moment M. Therefore, γ
According to the present embodiment in which t, φ, δt, and T are used as input signals to each neural network, the relationship between the acceleration / deceleration request and the steering request expressed by the basic motion equation of the vehicle and the motion state of the vehicle. Can be suitably reflected in the running stability control, and more accurate control can be realized as compared with the related art. Further, as to the generation of the torque commands TR and TL using the control neural network 312, there is no application restriction depending on the state such as whether or not the vehicle body is turning. Therefore, even when the vehicle is traveling straight, it is possible to suppress changes in the slip angle β and the yaw rate γt due to a slip or the like.

As described above, any one of the four types of input signals can be omitted, or can be replaced with another amount. In that case,
In addition to the relatively low-sensitivity yaw rate γt, it is preferable to use an amount having a higher sensitivity as the input signal. That is, as in the present embodiment, if the slip angle β having a relatively high sensitivity is used in addition to the yaw rate γt having a relatively low sensitivity, for example, the target of the yaw rate disclosed in Japanese Patent Application Laid-Open No. 5-91607 is used. Compared to control, accurate and highly reliable control can be realized. Further, when the learning control unit 604 generates the teacher signals TRS and TLS, the yaw rate γt, the slip angle β, and the steering angle δt which are also used as input signals to the control neural network 312 and the update neural network 602 are used. By using the torque command T and the torque command T, it is possible to generate teacher signals TRS and TLS with relatively high reliability. For example, in accordance with the feedback control logic proposed by the present applicant in Japanese Patent Application No. 9-68571, the teacher signals TRS and T
When the LS is generated, the model based on the basic motion equation of the vehicle body can be reflected even during learning during traveling. Therefore, the updating neural network 602 learns based on the output of the control neural network 312, and Neural network for control 31 based on
The learning / organization update loop of updating the second organization does not become unstable.

[0058]

As described above, in the preferred embodiment of the present invention, the control neural network learns in advance based on the results of actual vehicle tests or simulation calculations, not the basic body motion equations of the vehicle body that model the vehicle body motion. In addition to generating commands to the motor based on the contents of the current state, an update neural network that learns while the vehicle is running based on acceleration / deceleration requests, vehicle body motion status, control neural network output, etc. is provided, and the control neural network is organized even while the vehicle is running. Updating can be performed, which solves the problems inherent in the modeling of vehicle body motion based on the basic equations of motion.In addition, the steering angle is too large or too small in the actual use environment, disturbance received after steering, control system delay, and various other factors. Can deal with various control error factors. As a result, control reliability, accuracy, responsiveness, and the like are improved.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a system configuration of an electric vehicle suitable for carrying out the present invention.

FIG. 2 is a cross-sectional view showing an example of the structure of a wheel-in motor, and a circle in the upper right in the figure is a partially enlarged view.

FIG. 3 is a diagram showing a braking force distribution in the system of FIG. 1;

FIG. 4 is a block diagram illustrating a functional configuration of a vehicle control unit.

FIG. 5 is a block diagram illustrating a functional configuration of a torque command provisional determination unit.

FIG. 6 is a diagram showing the contents of a powering torque map.

FIG. 7 is a diagram showing the contents of a regenerative torque map.

FIG. 8 is a block diagram illustrating a functional configuration of a switching control unit.

FIG. 9 is a block diagram showing a functional configuration of a TRC / ABS equivalent control unit.

FIG. 10 is a diagram showing the contents of a coefficient determination map.

FIG. 11 is a block diagram illustrating a functional configuration of a traveling stability control unit and a learning management unit.

FIG. 12 is a map showing functions of a torque limit calculation unit.

FIG. 13 is a diagram conceptually showing a functional configuration of a control and update neural network.

FIG. 14 is a flowchart showing a learning procedure executed by the control and update neural networks before the vehicle starts running.

FIG. 15 is a flowchart showing a procedure of command generation and tissue update executed by the control neural network during running of the vehicle.

FIG. 16 is a flowchart showing a procedure of learning executed by the updating neural network while the vehicle is running, and a procedure of transferring the result.

FIG. 17 is a flowchart illustrating a teacher signal generation procedure executed by the learning control unit while the vehicle is traveling.

FIG. 18 is a conceptual plan view defining a predetermined number and variables for describing the motion of the vehicle body.

FIG. 19 is a conceptual diagram that defines a predetermined number and variables for describing the rotational motion of the wheel.

[Explanation of symbols]

10RR right rear wheel, 10RL left rear wheel, 12R, 12L
Motor, 34R, 34L Inverter, 36R, 36
L motor control unit, 38 vehicle control unit, 40RR, 40
RL, 40FR, 40FL Wheel speed sensor, 42 accelerator sensor, 44 brake sensor, 46 shift position switch, 48 steering angle sensor, 50 yaw rate sensor, 54 lateral acceleration sensor, 100 torque command provisional determination unit, 200 TRC / ABS equivalent control unit , 300
Running stability control section, 302 yaw rate calculation section, 304
Lateral acceleration calculator, 306 steering angle calculator, 308 slip angular velocity calculator, 310 slip angle calculator, 312 control neural network, 314 torque limit calculator, 400 control switcher, 500 switch controller, 600
Learning management unit, 602 updating neural network, 604 learning control unit, TR, TL, T torque command, φ body slip angle, γt yaw rate, δt steering angle, Wij, Vkj synaptic coupling coefficient, θj, λk
Neuron threshold.

Claims (1)

[Claims] 1. A drive control device for controlling the running of an electric vehicle by individually giving a command relating to the output to each of a plurality of motors for individually driving each driving wheel of the electric vehicle. Prior to the start of traveling, the learning of the generation of the above-mentioned command corresponding to the acceleration / deceleration request and the vehicle body movement state based on the actual vehicle test or the simulation calculation result has been learned. A control neural network for generating and outputting commands, and learning of generation of the commands according to the acceleration / deceleration request and the vehicle body motion state based on the actual vehicle test or simulation calculation results before the vehicle starts running, and the input signal during the vehicle running. The learning is continued based on the output of the control neural network, and the organization of the control neural network is updated based on the result. An updating neural network, and a traveling learning management means for providing a teacher signal to the updating neural network during traveling of the vehicle based on the input signal and the output of the controlling neural network,
A drive control device comprising: