JP4875906B2 - Vehicle behavior estimation / prediction device and vehicle stabilization control system - Google Patents

Description

  The present invention relates to a vehicle behavior estimation / prediction device and a vehicle stabilization control system, and more particularly to a vehicle behavior estimation / prediction device and a vehicle stabilization control system for calculating an estimated value or a predicted value of an amount representing a vehicle behavior by a neural network calculation.

  Development of so-called by-wire technology, in which operations of a steering wheel, an accelerator pedal, a brake pedal, and the like are read by a sensor or the like and transmitted as an electric signal to a drive system to perform electronic control, has been developed. Along with the sophistication of vehicle control, the amount of vehicle behavior used for control, that is, yaw rate, lateral acceleration, slip angle, lateral force and longitudinal force applied to tires, roll rate, pitch rate, road surface It is very important to accurately estimate current values such as a friction coefficient.

  In addition, if it is possible to predict the values of these quantities after a certain time and predict future vehicle behavior, it is considered that more accurate behavior control will be possible, and if such vehicle behavior prediction is possible, It is expected that vehicle behavior stabilization control can be performed before the vehicle behavior falls into an unstable state, leading to the development of more advanced preventive safety technology.

  However, there are many cases in which these measuring devices that measure the amount of vehicle behavior are difficult to mount on an actual vehicle because they are, for example, unreliable, expensive, or large. Therefore, in the past, such an amount that is difficult to measure in practice based on the measured value of the amount that can be measured relatively easily and accurately, such as the wheel speed and the steering angle of the steering wheel, using a mathematical model of vehicle motion. The value was estimated or the value after a certain time was predicted.

  In recent years, a vehicle behavior estimation / prediction device that calculates an amount representing a vehicle behavior by calculation using a neural network instead of using a mathematical model has been developed (see, for example, Patent Documents 1 and 2). However, it is difficult to accurately represent the vehicle behavior when the vehicle running state is in a non-linear region such as a spin state or a drift state, regardless of the computation by the neural network or the mathematical model. is there.

Therefore, the output value is delayed by one to several cycles in order to reflect the past history of the output value in the output value from the current neural network for the purpose of accurately estimating and predicting the amount representing the vehicle behavior even in the non-linear region. A vehicle behavior estimation / prediction device using a recurrent neural network configured to be fed back and input again is known (see Patent Document 3).
JP-A-11-147460 JP-A-6-286630 JP 2004-249812 A

  When using neural networks to calculate estimated values or predicted values that represent vehicle behavior, it is expected that the accuracy will be improved if many other types of quantities that may affect fluctuations in the quantity are input. , Usually the accuracy goes up.

  However, when the input to the neural network becomes multidimensional, the variables to be optimized by a genetic algorithm or the like become enormous, and the learning ability of the neural network may be saturated and may not easily converge. In particular, when a past output value is re-input as in the case of a recurrent type neural network, the input dimension increases as the number of delay cycles increases, so that there is a high possibility that learning cannot be performed easily.

  On the other hand, simply reducing the number of vehicle behaviors that are input to the neural network will reduce the accuracy of the output value, making it impossible to output a valid value as the current value or predicted value of the vehicle behavior. Therefore, the reliability of the vehicle behavior estimation / prediction device decreases.

  The present invention has been made in view of the above circumstances, and by reducing the number of inputs to the neural network and accurately performing vehicle behavior by dimensionally compressing quantities representing a plurality of vehicle behaviors while reflecting their characteristics. It is an object of the present invention to provide a vehicle behavior estimation / prediction device capable of estimating and predicting an amount to be represented and a vehicle stabilization control system using the same.

In order to solve the above problem, the first invention provides:
In the vehicle behavior estimation and prediction device,
Feature quantity extraction means for dimensionally compressing data relating to quantities representing a plurality of vehicle behaviors measured by the measuring device using an hourglass neural network and extracting an output from the compression layer of the hourglass neural network as a feature quantity;
An arithmetic means configured by a neural network that outputs an estimated value of the amount representing the current vehicle behavior based on the extracted feature amount or a predicted value of the amount representing the vehicle behavior after a certain time , and
The neural network constituting the calculation means is a recurrent neural network having a delay element feedback,
The recurrent neural network having the delay element feedback includes the hourglass neural network for dimensionally compressing the delay elements that are the estimated values or the predicted values output at a plurality of current or past time points, and the hourglass the output from the compression layer type neural network, characterized that you have been configured to input to the input layer of the recurrent neural network.

According to a second aspect of the present invention, in the vehicle behavior estimation / prediction device according to the first aspect, the hourglass neural network for extracting the feature value, the recurrent neural network, and the hourglass neural network for dimensionally compressing the plurality of delay elements, The actual measurement data measured by the measurement device and the output data from the recurrent neural network are individually learned in advance based on evaluation values calculated as teacher data.

According to a third aspect of the present invention, in the vehicle behavior estimation / prediction device of the second aspect, the hourglass neural network that extracts the feature amount, the recurrent neural network, and the hourglass neural network that dimensionally compresses the plurality of delay elements 2, further comprising a switching means for selecting any one of them in order to perform learning individually while switching the learning timing.

According to a fourth aspect of the present invention, in the vehicle behavior estimation / prediction device according to the third aspect of the present invention, the switching means includes an average value, a variance, and a period for the actual measurement data measured by the measurement device and the output data from the recurrent neural network, respectively. When the difference between the average value, variance and period of both is normalized and compared, and the difference between the normalized average values is the largest, the hourglass neural network that extracts the feature amount, Selecting the recurrent neural network when the normalized variance difference is the largest, and selecting the hourglass neural network that dimensionally compresses the plurality of delay elements when the normalized difference is the largest. It is characterized by.

A fifth invention is a vehicle stabilization control system,
A vehicle behavior estimation / prediction device according to any one of the first to fourth inventions;
And a control device that performs vehicle stabilization control based on the estimated value or the predicted value output from the vehicle behavior estimation / prediction device.

  According to the first invention, in the hourglass neural network of the feature quantity extraction means, the handle angle and the like are input to the input layer, and after the dimensional compression in the compression layer, the handle angle and the like are restored from the output layer and output. Is done. Therefore, it can be said that the output from the compression layer is information reflecting all the features of the quantities representing a plurality of vehicle behaviors such as the input steering angle.

  Therefore, by extracting the information from the compression layer, extracting it as a feature value, and inputting it to the calculation means, an amount representing a plurality of vehicle behaviors such as a steering angle is input to the neural network of the calculation means with one input. The same effect can be obtained. Therefore, it is possible to obtain the same effect as inputting many other types of quantities that may affect fluctuations in estimated values and predicted values representing the vehicle behavior output by the neural network to the neural network. Thus, it is possible to improve the accuracy of the estimated value and the predicted value of the quantity representing the vehicle behavior output from the neural network.

  In addition, it is possible to reduce the number of data that is actually input to the neural network by inputting the number representing the vehicle behavior into a small number of feature quantities, and reducing the number of data that is actually input to the neural network. Can also be reduced. As a result, the number of variables to be optimized by the genetic algorithm, etc. is small, and it is easily converged in the learning of the neural network, the learning efficiency is improved and the learning speed is promoted, and the vehicle behavior output from the neural network is It is possible to further improve the accuracy of the estimated value and the predicted value of the amount to be expressed.

According to the first invention, to reflect the estimated value or the like output before before or several cycles estimate and the predicted value of the quantity representing the vehicle behavior to be outputted from the arithmetic means, the estimated value and the predicted output It is possible to prevent an unexpected value from being output as a value, and it is possible to provide continuity and causality to the output estimated value and predicted value. Therefore, it is possible to estimate and predict the amount representing the vehicle behavior with high accuracy, and to exhibit the effect of the invention more accurately.

Further , according to the first invention, it is possible to dimensionally compress the information reflecting the output characteristics of the past several cycles output in the past into one or a small number of information and input it to the recurrent neural network. Therefore, the same effect as described above can be obtained for the delay element, and the effect of the invention can be more effectively exhibited.

According to the second invention, an hourglass neural network that extracts feature values, a recurrent neural network, and an hourglass neural network that dimensionally compresses a plurality of delay elements are learned using measured data and output data as teacher data, respectively. Thus, the convergence of learning of each neural network is improved as compared with the case where the entire neural network of the vehicle behavior estimation / prediction device is learned collectively. Therefore, the learning efficiency is further improved and the learning speed is promoted, and it is possible to further improve the accuracy of the estimated value and the predicted value representing the vehicle behavior output from the neural network. Is more effective.

According to the third invention, the hourglass neural network that extracts the feature value by the switching means, the recurrent neural network, and the hourglass neural network that dimensionally compresses the plurality of delay elements are selected and individually switched while switching the learning timing. Since learning is performed, each neural network can be efficiently learned, and the effects of the above-described inventions can be exhibited more efficiently.

According to the fourth invention, based on the average value, variance, and period of each waveform of the estimated value and predicted value output data representing the vehicle behavior output from the computing means and the teacher data used for learning. The learning timing is switched by switching means. If the difference between the average values of each waveform is large, the learning of the hourglass neural network that extracts the feature amount is bad, and if the variance of each waveform is large, the learning of the recurrent neural network is bad. If the difference in waveform period is large, the learning of an hourglass neural network that dimensionally compresses a plurality of delay elements is poor.

  Therefore, it is possible to appropriately select the neural network to be learned by switching the learning timing based on the average value, variance, and period of each waveform of the output data and the teacher data, and the vehicle behavior estimation / prediction device outputs Since the accuracy of the quantity representing the vehicle behavior to be performed can be improved efficiently and accurately, the effects of the inventions described above are more effectively exhibited.

According to the fifth aspect of the present invention, the current behavior of the vehicle is estimated based on the highly accurate estimated value and the predicted value transmitted from the vehicle behavior estimation / prediction device to the control device of the vehicle stabilization system, and the future behavior is determined. It becomes possible to predict. Therefore, it is possible to effectively control the factor that destabilizes the vehicle behavior using the highly accurate estimated value and predicted value, or to correct the factor that destabilizes the vehicle behavior and Thus, it is possible to effectively control in advance so as not to occur.

  In addition, the calculation means of the vehicle behavior estimation / prediction device preliminarily learns the driving data in the state where the road surface is dry and the disturbance is small as the teacher data, thereby estimating the amount representing the vehicle behavior in the low disturbance state By outputting values and predicted values and comparing them with actual measured values, it is possible to grasp the disturbances that may or may occur in the vehicle and accurately control the stabilization of vehicle behavior Become.

  Embodiments of a vehicle behavior estimation / prediction device and a vehicle stabilization control system according to the present invention will be described below with reference to the drawings.

  First, the vehicle behavior estimation prediction apparatus according to the present embodiment will be described. The vehicle behavior estimation / prediction device 1 according to the present embodiment is configured by a computer in which a CPU, a ROM, a RAM, an input / output interface and the like (not shown) are connected by a bus.

  FIG. 1 is a block diagram illustrating a configuration of a vehicle behavior estimation prediction apparatus according to the present embodiment. The vehicle behavior estimation / prediction device 1 includes a feature amount extraction unit 2, a calculation unit 3, and an optimization unit 4.

  In the present embodiment, the vehicle behavior estimation / prediction device 1 includes a steering angle sensor S1, a steering reaction force sensor S2, a wheel speed sensor S3, a lateral acceleration sensor S4, a longitudinal acceleration sensor S5, an accelerator opening sensor S6, a temperature sensor S7, A measuring device S such as a wind speed sensor S8 is connected, and from each sensor, a steering wheel angle δ, a steering reaction force St, a wheel speed V, a vehicle body lateral acceleration Gy, a vehicle body longitudinal acceleration Gx, an accelerator opening A, an outside air temperature Te, The outside air wind speed W is input.

The feature amount extraction means 2 is inputted with data of the steering wheel angle δ, the steering reaction force St, the wheel speed V, the vehicle body lateral acceleration Gy, the vehicle body longitudinal acceleration Gx, and the accelerator opening A. Further, with respect to the steering wheel angle δ, the differential value and the second-order differential value, that is, the steering wheel angular velocity dδ and the steering wheel angular acceleration d ² δ are calculated by the differentiators 5 and 6, and these values are also input to the feature amount extraction means 2. It has become.

The feature quantity extraction means 2 is composed of one or a plurality of hourglass neural networks inSNN as shown in FIG. In the present embodiment, the steering wheel angle δ, the steering wheel angular velocity dδ, the steering wheel angular acceleration d ² δ, and the steering reaction force St are input to one hourglass neural network. The vehicle body lateral acceleration Gy, the vehicle body longitudinal acceleration Gx, and the accelerator opening A are input to another hourglass neural network.

  The hourglass neural network inSNN used in the present embodiment includes an input layer L1 composed of nodes N1 to N4, an input side intermediate layer L2 composed of nodes N5 to N7, a compression layer L3 composed of a node N8, a node The output side intermediate layer L4 includes N9 to N11 and the output layer L5 includes nodes N12 to N15.

  Note that the nodes constituting the compressed layer L3 can be constituted by a plurality of nodes as shown in FIG. 2B, for example. In this case, the number of nodes constituting the compression layer L3 is set to be smaller than the number of nodes in the input layer L1, that is, the number of input data.

  In the hourglass neural network inSNN, nodes are arranged so that the input side and the output side are symmetric with respect to the compression layer L3. The number of nodes constituting the input-side intermediate layer L2 and the output-side intermediate layer L4 is set as appropriate, or they may not be provided.

  The hourglass neural network inSNN is configured such that the same data as the input data to the input layer L1 is output from the output layer L5 by learning described later. That is, the data is dimensionally compressed by the flow from the input layer L1 to the compression layer L3 and restored by the flow from the compression layer L3 to the output layer L5.

  In other words, the output from the compression layer L3 to the output-side intermediate layer L4 is input data such that the input data is restored by the flow from the compression layer L3 to the output layer L5, that is, a quantity characteristic representing a plurality of input vehicle behaviors. The amount is all reflected. Therefore, the output from the compression layer L3 can be used as a feature amount that represents a feature of an amount that represents a plurality of input vehicle behaviors.

  Therefore, in the present invention, the feature quantity of the quantity representing the plurality of vehicle behaviors inputted by the dimension compression of the data related to the quantity representing the plurality of vehicle behaviors inputted by the hourglass neural network inSNN and the output from the compression layer L3. And the output from the compression layer L3, that is, the feature value is output to the calculation means 3.

  In the present embodiment, each output from the output layer L5 of the hourglass neural network inSNN is not used except during learning described later.

  In the hourglass neural network inSNN shown in FIG. 2 (A), an hourglass type in which one type of feature value f1 is output from the compression layer L3 consisting of one node to the computing means 3 and another wheel speed V is input. Another feature amount f2 is output to the calculation means 3 from the compression layer of the neural network inSNN. In the hourglass neural network inSNN shown in FIG. 2B, two types of feature quantities are output from the compression layer composed of two nodes, and either or both of them are output and sent to the computing means 3. .

Each node Ni constituting the input layer L1, input side intermediate layer L2, compression layer L3, output side intermediate layer L4 and output layer L5 of the hourglass neural network inSNN has one output for the input Xi according to a predetermined transfer function. Yi is output. In this embodiment, a sigmoid function represented by the following equation (1) is used as a transfer function. In the following equation (1), θi is a threshold value for each node Ni.
Yi = 1 / (1 + exp (− (Xi−θi))) (1)

In the input layer L1, the input Xi in the expression (1) is a normalized value of each of the steering wheel angle δ, the steering wheel angular velocity dδ, the steering wheel angular acceleration d ² δ, and the steering reaction force St input to the nodes N1 to N4. It is. In the other hourglass neural network inSNN, the wheel speed V, the vehicle body lateral acceleration Gy, the vehicle body longitudinal acceleration Gx, and the accelerator opening A are normalized values.

In the input side intermediate layer L2, the compression layer L3, the output side intermediate layer L4, and the output layer L5, the input Xi in the equation (1) is a sum obtained by weighting the output Yj of each upstream node Nj. . That is, when the connection weight coefficient between the upstream node Nj and the downstream node Ni is wij, the input Xi to the node Ni is
Xi = Σwij × Yj (2)
Is required.

As the transfer function, a function other than the sigmoid function shown in the equation (1) can be used. For example, it is possible to use functions such as hysteresis, limiter, binarization, dead zone, time delay of delay circuit Z- ^1, etc., peak hold, differentiation, integration, maximum value, etc., which are appropriately selected and used.

  If there is a large amount of data to be dimensionally compressed, the data output from the compression layers of a plurality of hourglass neural networks are input again to another hourglass neural network and further dimensionally compressed. It is also possible to send the output as a feature quantity to the calculation means 3.

  As shown in FIG. 1, the calculation means 3 corresponds to functions required for the vehicle behavior estimation prediction apparatus 1 such as a slip angle calculation means 31, a yaw rate calculation means 32, a roll rate calculation means 33, a pitch rate calculation means 34, and the like. It is composed of one or a plurality of arithmetic means. Therefore, in addition to this, for example, it is possible to provide a calculation means for calculating a lateral force, a longitudinal force applied to the tire, a road surface friction coefficient, and the like.

  Each of the calculation means constituting the calculation means 3 is supplied with the extracted feature value transmitted from the feature value extraction means 2 and the measurement value transmitted from each measurement device S. ing. In the present embodiment, the measured values of the outside air temperature Te and the outside air wind speed W transmitted from the temperature sensor S7 and the wind speed sensor S8 are directly input to the calculation means 3.

  In addition, each computing means calculates an estimated value representing the current vehicle behavior or a predicted value representing the vehicle behavior after a certain time, or both, based on the respective feature values and measured values. ing. Note that whether an estimated value or a predicted value is output from each computing means, or whether both can be output, depends on the structure of the output layer of the recurrent neural network RNN, which will be described later, and what is output. It depends on how.

  In the present embodiment, each computing means constituting the computing means 3 is constituted by a hierarchical recurrent neural network RNN having delay element feedback as shown in FIG.

  In the following description, it is assumed that the recurrent neural network RNN having delay element feedback including the hourglass neural network retSNN outputs the current slip angle estimation value βe. The same applies to the case where the predicted value βe is output or the case where the estimated value or the predicted value of an amount representing another vehicle behavior is output.

  The main part of the recurrent neural network RNN used in this embodiment includes an input layer L6 composed of nodes N16 to N20, a first intermediate layer L7 composed of nodes N21 to N23, and nodes N24 to N27. The output layer L9 includes a second intermediate layer L8 and a node N28.

  Note that only one intermediate layer may be provided instead of two, or three or more layers may be provided. Further, the number of nodes constituting the intermediate layers L7 and L8 and the output layer L9 is set as appropriate. When the output layer L9 is composed of a plurality of nodes, it is determined as appropriate from which node the estimated value or the predicted value is output and the output from which node is transmitted to the hourglass neural network retSNN described later.

  Each node Ni constituting the input layer L6, the first intermediate layer L7, the second intermediate layer L8, and the output layer L9 of the recurrent neural network RNN is 1 for the input Xi according to the sigmoid function expressed by the above equation (1). The outputs Yi are output, and the inputs Xi in the first intermediate layer L7, the second intermediate layer L8, and the output layer L9 weight the outputs Yj of the respective upstream nodes Nj according to the equation (2). The sum is the same as in the case of the hourglass neural network inSNN described above. It is also possible to use a function other than the sigmoid function as the transfer function.

  In the nodes N16 to N19 constituting the input layer L6 of the recurrent type neural network, the normalized values of the feature values f1 and f2 and the measured values of the outside air temperature Te and the outside air wind speed W transmitted from the feature amount extracting unit 2 are provided. Is entered.

  In addition, the estimated value βe output from the recurrent neural network at a plurality of current or past time points is fed back as a delay element and dimension-compressed and input to another node N20 constituting the input L6. It has become.

  For the dimension compression of the delay element, an hourglass neural network retSNN similar to the hourglass neural network inSNN that constitutes the feature quantity extraction means 2 is used. In this embodiment, the hourglass neural network retSNN has the same configuration as the hourglass neural network inSNN, and includes an input layer L10, an input side intermediate layer L11, a compression layer L12, an output side intermediate layer L13, and an output layer L14. Has been.

The node N29 constituting the input layer L10 of the hourglass neural network retSNN receives the estimated value βe that is currently output as a delay element, and each of the nodes N30 to N32 has one cycle in the delay circuit Z- ¹ in synchronization therewith. Each past estimated value βe delayed by one is input as a delay element.

  In the hourglass neural network retSNN, the same data as the input data to the input layer L10 is output from the output layer L14 by learning, and the dimensionally compressed information is output from the compression layer L12. Yes. Information output from the compression layer L12 is input to the node N20 of the input layer L6 of the recurrent neural network RNN.

  This information is an amount that reflects the past history of the estimated slip angle βe that is the output of the recurrent neural network RNN, that is, a temporal change, that is, an amount that can be said to be a feature amount fd of the temporal change of the output value. is there.

  In this embodiment, the case where the current output value and the continuous output value one cycle before the current output value are input to the hourglass neural network retSNN has been described. The time interval of values and the number of values used are appropriately determined. Further, the number of nodes in the input layer L10 and the same number of nodes in the output layer L14 of the hourglass neural network retSNN are determined depending on how many estimated values βe in the current and past are fed back.

  At that time, if the number of input data to the input layer L10 increases, the data output from the compression layers of the plurality of hourglass neural networks are input again to another hourglass neural network and further dimensionally compressed, It is also possible to configure so that the output from the compression layer is input to the input layer L6 of the recurrent neural network RNN.

  Further, in the present embodiment, as in the case of the hourglass neural network inSNN, each output from the output layer L14 of the hourglass neural network retSNN is not utilized except during learning described later.

  The optimizing means 4 includes the hourglass neural network inSNN constituting the feature quantity extracting means 2 and the recurrent neural network RNN and hourglass neural network retSNN in each computing means constituting the computing means 3 for each node Ni. The threshold θi and the coupling weight coefficient wij are automatically optimized in advance by learning based on the learning rule of the genetic algorithm.

  In the following, a case where learning is performed for the slip angle calculation means 31 that outputs the current slip angle estimation value βe as the calculation means 3 will be described. However, when the predicted slip angle βe after a certain time is output, The same applies to the case where an estimated value or predicted value of an amount representing another vehicle behavior is output. In addition, the case where the transfer function of each neural network node is a sigmoid function will be described. However, when the transfer function is composed of functions other than the sigmoid function, the thresholds, coefficients, constants, etc. used for the function are the same. To be optimized.

  As shown in FIG. 4, the optimization unit 4 includes a data storage unit 7, an evaluation unit 8, and a switching unit 9.

  In the data storage means 7, time-series measured data D used for learning is input and stored. The actual measurement data D is data common to the respective calculation means, and the steering wheel angle δ, the steering reaction force St, the wheel speed V, the vehicle body lateral acceleration Gy, the vehicle body longitudinal acceleration Gx, and the accelerator opening that are input to the vehicle behavior estimation / prediction device 1. A, measured values for a predetermined period of time, outside air temperature Te, and outside wind speed W, and are also used as teacher data.

  Further, the data storage means 7 is input with time-series data of the slip angle βre measured by the slip angle sensor at the same timing as the data D as the teacher data T used for the learning of the slip angle calculation means 31. And saved. Note that time-series measured values representing quantities representing vehicle behaviors used as teacher data for learning of each computing means constituting the computing means 3 are input and stored in the data storage means 7 in advance.

  As shown in FIG. 5, the evaluation unit 8 includes an initial individual generation unit 10, an evaluation value calculation unit 11, an end determination unit 12, a parent selection unit 13, a crossover unit 14, and a mutation unit 15. ing. The evaluation unit 8 includes a memory (not shown).

  First, learning of the hourglass neural network inSNN of the feature quantity extraction unit 2 will be described.

  In response to the optimization start instruction, the initial individual generation unit 10 preliminarily selects an individual including a gene with a necessary number of thresholds θi and a coupling weight coefficient wij for a sigmoid function of each node constituting the hourglass neural network inSNN. A set number of individuals is generated to form a first generation individual group.

  One individual of the genotype is represented by the threshold value θ1 of the node N1, the coupling weight coefficients w1,5, w1,6, w1,7, and the node for each of the nodes N1 to N15 of the hourglass neural network inSNN in FIG. Are formed in a form including a numerical sequence in which the threshold θi of each node and the coupling weight coefficient wij are arranged in one column in the order of the N2 threshold θ2, the coupling weight coefficient w2, 5,. At the time of generating the first generation individual group, the threshold value θi and the coupling weight coefficient wij of each node of each individual are set at random. Note that the configuration of the individual is not necessarily limited to such a configuration, and is appropriately determined from the viewpoint of the ease of the following calculation.

  When there are two hourglass neural networks inSNN constituting the feature amount extraction unit 2 as in the present embodiment, the initial individual generation unit 10 generates two first-generation individual groups independently. The evaluation means 8 is configured to perform the following evolution process independently in parallel in each of the two groups.

Hereinafter, learning of the hourglass neural network inSNN to which the steering wheel angle δ, the steering wheel angular velocity dδ, the steering wheel angular acceleration d ² δ, and the steering reaction force St are input will be described, but the wheel speed V, the vehicle body lateral acceleration Gy, and the vehicle body longitudinal acceleration Gx are described. The hourglass neural network inSNN to which the accelerator opening A is input is also learned in the same manner and simultaneously.

  The initial individual generating means 10 generates, for example, 100 individuals having the above-mentioned configuration per group and transmits them to the evaluation value calculating means 11.

  The evaluation value calculation means 11 selects one individual from the individual group, transmits the threshold θi of each node included in the individual and the coupling weight coefficient wij to the hourglass neural network inSNN, and the hourglass neural network inSNN. Try to set up and try.

When the setup is completed, the evaluation value calculation unit 11 reads the measured data for a predetermined time of the steering wheel angle δ and the steering reaction force St input to the data storage unit 7 in time series. Further, the steering wheel angle δ is subjected to differentiation and second-order differentiation using the differentiators 5 and 6 of the apparatus 1 to calculate the steering wheel angular velocity dδ and the steering wheel angular acceleration d ² δ. Then, the steering wheel angle δ, the steering wheel angular velocity dδ, the steering wheel angular acceleration d ² δ, and the steering reaction force St are represented by the nodes N1 to N4 of the input layer L1 of the hourglass neural network inSNN of the feature amount extraction means 2 shown in FIG. Are sequentially input in synchronization with each other.

When the evaluation value calculation unit 11 sequentially outputs the steering wheel angle δ, the steering wheel angular velocity dδ, the steering wheel angular acceleration d ² δ, and the steering reaction force St from the nodes N12 to N15 of the output layer L5 according to the input to the input layer L1. The evaluation value Q of each individual is calculated according to the following equation (3).

Here, Si is the steering wheel angle δ, the steering wheel angular velocity dδ, the steering wheel angular acceleration d ² δ, and the steering reaction force St output from the nodes N12 to N15, and Ti is the steering wheel angle δ and the steering wheel that are input to the nodes N1 to N4. Angular velocity dδ, steering wheel angular acceleration d ² δ, and steering reaction force St. Further, tc is a time required for one cycle, and N is a predetermined total number of cycles of the actual measurement data D for a predetermined time input from the data storage unit 7 to the evaluation value calculation unit 11.

  In this case, the actual measurement data D is not only the actual measurement value input but also teacher data. As can be seen from the equation (3), the smaller the evaluation value Q, the higher the degree of approximation between the input actual measurement data D and the output data.

  The evaluation value calculation means 11 performs this trial for all the individuals in the individual group, and calculates the evaluation value Q for each individual. The evaluation value calculation unit 11 transmits the individual and the evaluation value Q to the end determination unit 12.

  The end determination means 12 determines whether or not the set end determination condition is satisfied. When it is determined that the termination determination condition is satisfied, the threshold θi and the coupling weight coefficient wij of each node of the hourglass neural network inSNN included in the individual having the best evaluation value Q in the individual group are characterized. It is sent to the quantity extraction means 2 to set up an hourglass neural network, and at the same time, all individuals of the individual group are stored in a memory. Further, when it is determined that the end determination condition is not satisfied, the end determination unit 12 transmits the individual and the evaluation value Q to the parent selection unit 13.

  In the present embodiment, the end determination condition is that the set number of generations is reached, and the evolution process in the evaluation means 8 ends when the set number of generations is reached. However, it is also possible to set the end determination condition so as to end when an individual having an evaluation value Q that is better than the set evaluation value, that is, smaller than the set evaluation value, appears. It is also possible to set to end when the processing time is reached.

  The parent selection means 13 of this embodiment is configured to perform parent selection by elite storage. That is, the parent selection means 13 stores and isolates a predetermined number of individuals in the memory 16 in ascending order of the evaluation value Q, and simultaneously creates a copy of the isolated individual, and the copy and the stored individual. Individual individuals are selected by a method such as reverse roulette selection, expected value selection, ranking selection, tournament selection, and the like. The parent selection means 13 transmits the selected individual to the crossover means 14.

  In the crossover means 14, for each individual that should be a parent individual sent from the parent selection means 13, one-point crossover, multipoint crossover, uniform crossover, SPX (Simplex Crossover) crossover, UNDX (Unimodel Normal Distribution Crossover) crossover, etc. Crossing is performed with a predetermined probability by a normal method. The individual generated by the crossover is transmitted to the mutation means 15, and the mutation means 15 randomly changes the values of the threshold value θi and the coupling weight coefficient wij of each node constituting the individual at a constant rate. It is like that. In this way, a next generation population is generated.

  Each individual for which the processing in the crossover means 14 and the mutation means 15 has been completed is transmitted to the evaluation value calculation means 11, and a trial is performed again for each individual, and the evaluation value Q is calculated. . The individual for which the evaluation value Q is calculated is transmitted to the end determination unit 12, and the end determination unit 12 performs end determination together with the individual one generation before stored in the memory 16. This optimization process is repeated until the end condition is satisfied.

  The evaluation unit 8 learns the hourglass neural network inSNN of the feature amount extraction unit 2 as described above. Further, when learning the hourglass neural network inSNN again in the learning of the series of hourglass neural networks inSNN, retSNN and recurrent neural network RNN, the population for the hourglass neural network inSNN stored in the memory is used. Each individual is read from the memory and learning is continued.

  The evaluation means 8 learns the hourglass neural network retSNN in the arithmetic means 3 shown in FIG. 3 in the same manner as the learning of the hourglass neural network inSNN.

  However, a new individual group is generated separately from the individual group used in the hourglass neural network inSNN, and the hourglass neural network retSNN is trained. To learn independently.

  Further, when the termination determination condition is satisfied, the hourglass neural network of the computing means that has learned the threshold value θi and the coupling weight coefficient wij of each node included in the individual having the best evaluation value Q in the individual group. At the same time as sending it to the network retSNN for setup, all individuals of the individual group for the hourglass neural network retSNN are stored in the memory.

  Further, when learning the hourglass neural network retSNN again in the learning of a series of hourglass neural networks inSNN, retSNN, and recurrent neural network RNN, for the hourglass neural network retSNN stored in the memory. Each individual of the individual group is read from the memory, and learning is continued.

  In the learning of the hourglass neural network retSNN, the input data as the teacher data input to the nodes N29 to N32 of the input layer L10 is recurrent using the hourglass neural network retSNN as shown in FIG. Current estimated value βet which is output data from the neural network RNN and past estimated values βet-1, βet-2, βet-3 delayed by one cycle.

Therefore, the evaluation value calculation means 11 of the evaluation means 8 determines the steering wheel angle δ, the steering reaction force St, the wheel speed V, the vehicle body lateral acceleration Gy, and the vehicle body longitudinal acceleration Gx that are input to the data storage device 7 during each individual trial. Then, data of the accelerator opening A is read in time series, and the steering wheel angle δ is extracted by using the differentiators 5 and 6 to calculate the steering wheel angular velocity dδ and the steering wheel angular acceleration d ² δ using the differentiators 5 and 6. Transmit to means 2. At the same time, data on the outside air temperature Te and the outside air wind speed W is read from the data storage means 7 and transmitted to the computing means 3.

  Then, the current estimated value βet output from the recurrent neural network RNN and the past estimated values βet-1, βet-2, βet-3 delayed by one cycle are used as the nodes of the input layer L10 of the hourglass neural network retSNN. N29 to N32 are sequentially input, and the current estimated value βet and the past estimated values βet-1, βet-2, βet-3 are sequentially output from the nodes N40 to N43 of the output layer L14.

  Then, the evaluation value Q of each individual is calculated according to the equation (3). In this case, Si in the equation (3) is the estimated values βet, βet-1, βet-2, βet-3 output from the nodes N40 to N43, and Ti is the estimated value βet input to the nodes N29 to N32. , Βet-1, βet-2, and βet-3.

  In the present embodiment, also in the learning of the hourglass neural network retSNN, the end determination condition reaches the set number of generations. It is possible to set other conditions.

Also in the learning of the recurrent neural network RNN shown in FIG. 3, the evaluation value calculation means 11 of the evaluation means 8 performs the steering angle δ, the steering reaction force St, Data of wheel speed V, vehicle body lateral acceleration Gy, vehicle body longitudinal acceleration Gx, and accelerator opening A are read in time series, and the steering wheel angle δ is determined from the steering wheel angular velocity dδ by differentiation using the differentiators 5 and 6 and second-order differentiation. The steering wheel angular acceleration d ² δ is calculated and transmitted to the feature quantity extraction unit 2, and the data of the outside air temperature Te and the outside air wind speed W are read from the data storage unit 7 and transmitted to the calculation unit 3.

  In the case of the recurrent type neural network RNN, the data stored as the teacher data T in the data storage means 7, that is, in the case of the slip angle calculation means 31, the slip actually measured at the same timing as the handle angle δ and the like. The data of the angle βre is read out, and the evaluation value Q of each individual is calculated by substituting the data and the estimated value βet output from the recurrent neural network RNN into the equation (3).

  Accordingly, in this case, Si in the equation (3) is time series data of the estimated value βet output from the node N28, and Ti is an actual measured value βre of the slip angle β read from the data storage means 7 as teacher data. It is time series data. In this case, i = 1 since there is only one type of data.

  Individuals used for learning of a specific recurrent neural network RNN are individuals generated independently of individuals used for learning of other arithmetic means and other hourglass neural networks inSNN and retSNN. Is stored in the memory, and when learning the recurrent neural network RNN again in a series of learning, each individual of the individual group for the recurrent neural network RNN stored in the memory is read from the memory, The subsequent learning is the same as in the learning of the hourglass neural network inSNN and retSNN.

  In the present embodiment, the end determination condition also reaches the set number of generations in learning of the recurrent neural network RNN. It is possible to set other conditions.

  The switching means 9 selects any one neural network every time learning of one neural network is completed in order to perform learning of the hourglass neural network inSNN, retSNN and recurrent neural network RNN while switching learning timing. It is means to do.

  As a reference for switching the learning timing in the switching means 9, for example, it is possible to simply perform learning in the order of the hourglass neural network inSNN, the recurrent type neural network RNN, and the hourglass neural network retSNN.

  In this embodiment, as a reference for switching the learning timing in the switching means 9, a neural network to be learned is selected according to the following criteria.

  That is, the handle angle δ or the like is input to the apparatus 1 to output the estimated value βe from the recurrent neural network RNN of the computing means, and the estimated value βe and the actual measured value βre of the slip angle that is the teacher data Calculate the mean, variance, and period for the waveform. When the average value, variance, and period difference between the two are normalized and compared, the hourglass neural network inSNN is used when the difference between the two is the average value. When the recurrent type neural network RNN has a period, the hourglass type neural network retSNN is selected.

  Specifically, the switching means 9 first selects the individual having the best evaluation value Q from among the individual groups of the hourglass neural network inSNN, retSNN, and recurrent neural network RNN stored in the memory at the present stage. The threshold θi and the coupling weight coefficient wij of each included node are transmitted to each neural network and set up.

Then, the data of the steering wheel angle δ, the steering reaction force St, the wheel speed V, the vehicle body lateral acceleration Gy, the vehicle body longitudinal acceleration Gx, and the accelerator opening A are read out in time series from the data storage means 7. 5 and 6 are used to calculate the steering wheel angular velocity dδ and the steering wheel angular acceleration d ² δ by differentiation or second-order differentiation and transmit them to the feature amount extraction means 2, read out the data of the outside air temperature Te and the outside air wind speed W, and the computing means 3. Is transmitted to the slip angle calculating means 31.

  The time series data of the estimated slip angle βe output from the recurrent neural network RNN is divided into time series data βei for each predetermined time interval, and the average value Mei, variance σei, period Fei for each time series data βei. Are calculated respectively. In this case, the period Fei is a reciprocal of the frequency corresponding to the maximum peak of the frequency spectrum obtained by performing fast Fourier transform (FFT) on the time series data βei to obtain a frequency spectrum.

  Further, the switching means 9 reads the slip angle βre as the teacher data measured at the same timing as the handle angle δ etc. from the data storage means 7, and similarly divides it into time series data βrei for every predetermined time interval, For each time series data βrei, an average value Mrei, variance σrei, and period Frei are calculated.

  Then, the difference between the average value Mei and the average value Mrei is calculated and added together. Differences are similarly calculated and added to the variance σeei and variance σrei, and the cycle Fei and cycle Frei. Then, the average value difference, variance difference, and period difference added together are normalized, and the respective ratios are calculated and the one with the highest ratio is selected as the learning target.

  In the present embodiment, the average value is associated with the hourglass neural network inSNN, the variance is associated with the recurrent neural network RNN, and the period is associated with the hourglass neural network retSNN.

  This is because, if the learning of the hourglass neural network inSNN of the feature quantity extraction means 2 is poor, a difference occurs between the average value of the output data βe and the average value of the teacher data βre, and the recurrent neural network of the calculation means 3 If the learning state of the RNN is poor, a difference occurs between the variance of the output data βe and the variance of the teacher data βre. If the learning state of the hourglass neural network retSNN is poor, the cycle of the output data βe and the cycle of the teacher data βre Based on the finding that there is a difference.

  When the evaluation means 8 finishes learning the set number of generations, the switching means 9 calculates the average value difference, variance difference, and period difference in the above manner, for example, (average value difference) :( distribution) Difference): (period difference) is 0.1: 0.4: 0.5, the hourglass neural network retSNN is selected as a learning target and the evaluation means 8 receives the hourglass neural network retSNN. Learn to do it.

  In the present embodiment, the switching is terminated when the number of switching by the switching unit 9 reaches a predetermined number, and the optimization process in the optimization unit 4 is terminated. As other termination conditions, for example, when the individual best evaluation values Q for the hourglass neural network inSNN and retSNN and the recurrent neural network RNN exceed all the set evaluation values, respectively. It is also possible to configure so as to end the process.

  Next, a vehicle stabilization control system using the vehicle behavior estimation / prediction device 1 according to the present embodiment will be described.

  As shown in FIG. 6, the vehicle stabilization control system 17 according to the present embodiment includes the vehicle behavior estimation / prediction device 1, a control device 18, and various actuators 19 that adjust the accelerator opening and the like. . Specific examples of the various actuators 19 include a four-wheel independent brake, an electronically controlled throttle for controlling the accelerator opening, and a system having a steering servo function such as steering control by a by-wire.

  When the steering angle δ measured by the measuring device S such as the steering angle sensor S1 of the steering wheel is input as described above, the vehicle behavior estimation / prediction device 1 receives the current as described above with respect to the control device 18. The estimated slip angle value βe, the estimated yaw rate value γe, the predicted slip angle value βe after a certain time, the predicted yaw rate value γe, and the like are output.

  Further, the calculation means 3 such as the slip angle calculation means 31 and the yaw rate calculation means 32 of the vehicle behavior estimation / prediction device 1 according to the present embodiment learns driving data in a state where the road surface is dry and the disturbance is small as teacher data. Has been done.

  The control device 18 is configured by a computer in which a CPU, ROM, RAM, input / output interface and the like (not shown) are connected to a bus. The control device 18 includes a slip angle and a yaw rate output from the vehicle behavior estimation prediction device 1. Etc. and predicted values βe and γe are input. In addition, in this embodiment, the steering angle δ, and the actual slip angle βre and the actual yaw rate γre measured by a slip angle sensor and a yaw rate sensor (not shown) are input.

  The control device 18 compares the current estimated values βe and γe of the slip angle and yaw rate at the steering angle δ of the input steering wheel with the measured values βre and γre measured by the sensor to determine whether the vehicle is in a spin state or a drift state. Alternatively, it is determined whether the state is an understeer state or an oversteer state.

  Various actuators 19 and the like are operated in accordance with a control signal transmitted from the control device 18.

  Next, the operation of the vehicle behavior estimation / prediction device 1 according to the present embodiment and the vehicle stabilization control system 17 using the same will be described.

  The hourglass neural network inSNN of the feature quantity extraction means 2 of the vehicle behavior estimation and prediction device 1 and the recurrent neural network RNN and hourglass neural network retSNN of each arithmetic means constituting the arithmetic means 3 are converted into the steering angle δ by the optimization means 4. Are previously learned and optimized based on the actual measurement data D and the teacher data T. Then, each node is set up by the optimized threshold value θi of each node and the coupling weight coefficient wij.

Measurement values such as steering angle δ, steering reaction force St, and wheel speed V measured from the steering angle sensor S1, the steering reaction sensor S2, the wheel speed sensor S3, and the like are synchronized with the vehicle behavior estimation / prediction device 1. When input at every fixed sampling cycle, among the measured values, the handle angle δ calculates the handle angular velocity dδ and the handle angular acceleration d ² δ by the differentiators 5 and 6, and the handle angle δ, the handle angular velocity dδ, The steering wheel angular acceleration d ² δ, the steering reaction force St, the wheel speed V, the vehicle body lateral acceleration Gy, the vehicle body longitudinal acceleration Gx, and the accelerator opening A are input to the feature amount extraction means 2. Further, the outside air temperature Te and the outside air wind speed W are input to the calculation means 3.

In the feature quantity extracting means 2, normalized values of the steering wheel angle δ, the steering wheel angular velocity dδ, the steering wheel angular acceleration d ² δ, and the steering reaction force St are input to the nodes N1 to N4 of the input layer L1 of one hourglass neural network inSNN. Then, the feature amount f1 is extracted from the node N8 of the compression layer L3 after being dimensionally compressed, and is output to the calculation means 3. Further, the normalized values of the wheel speed V, the vehicle body lateral acceleration Gy, the vehicle body longitudinal acceleration Gx, and the accelerator opening A are respectively input to the nodes N1 to N4 of the input layer L1 of the other hourglass neural network inSNN, and the dimensions are compressed. A feature value f2 is extracted from the node N8 of the compression layer L3 and output to the computing means 3.

  Each calculation means constituting the calculation means 3 has the normalized values of the feature quantities f1 and f2 transmitted from the feature quantity extraction means 2 and the normalized values of the outside air temperature Te and the outside air wind speed W that are directly input. Entered. And each calculating means calculates the estimated value and predicted value of the quantity showing the vehicle behavior which should each calculate simultaneously. It is also possible to select a calculation means to perform a calculation from the calculation means and perform the calculation.

  The slip angle calculating means 31 will be described as an example. The normal values of the characteristic quantities f1, f2, the outside air temperature Te, and the outside air speed W are respectively applied to the nodes N16 to N19 of the input layer L6 of the recurrent neural network RNN constituting the slip angle calculating means 31. The normalized value of the feature value fd of the temporal change of the output value βe is input from the compressed layer L12 of the hourglass neural network retRNN to the node N20.

  Each input value is subjected to a neural network calculation process, and an output value βe representing an estimated value or predicted value of the slip angle is output from the node N28 of the output layer L9 of the recurrent neural network RNN.

The output value βe is normalized through the delay circuit Z ⁻¹ and is input to the nodes N29 to N32 of the input layer L10 of the hourglass neural network retSNN, and is dimensionally compressed, and the feature value fd is output from the node N36 of the compression layer L12. Extracted and output, and input to the node N20 of the input layer L6 of the recurrent neural network RNN.

  The vehicle stabilization control system 17 receives, for example, the estimated slip angle βe and the estimated yaw rate γe from the vehicle behavior estimation / prediction device 1, and the steering angle δ measured by the measuring device S such as the steering angle sensor S1. And the actual slip angle βre and the actual yaw rate γre are input from the slip angle sensor and the yaw rate sensor.

  When these values are input, the control device 18 compares the estimated values βe and γe of the slip angle and yaw rate at the steering angle δ of the steering wheel with the measured values βre and γre by the sensor, for example, It is determined whether the vehicle is in a spin state, a drift state, an understeer state, or an oversteer state, and a control signal is transmitted to various actuators 19 and the like to perform vehicle stabilization control.

  Specifically, first, the control device 18 is in a state in which the estimated value γe of the yaw rate and the actual yaw rate γre are within a certain error range and are substantially equal, and the actual slip angle βre compared to the estimated value βe of the slip angle. Is greater than the set threshold, the road surface friction coefficient decreases and the vehicle is slipping on the road surface, and it is determined that the vehicle is spinning or is about to spin. A control signal is transmitted to various actuators 19 so as to prevent the generation.

  In this case, for example, the control device 18 transmits a signal so as to increase the brake pressure to a brake located outside the turning of the vehicle among the four-wheel independent brakes, and the accelerator is not opened to the electronic control throttle that controls the accelerator opening. A signal is transmitted to reduce the degree of rotation, and a system having a steering servo function transmits a signal to reduce the absolute value of the steering angle of the steering so as to prevent or generate spin. Eliminate spin.

  When the actual slip angle βre is the same or smaller than the estimated slip angle βe, the control device 18 does not perform any special control on the various actuators 19 and the like.

  Further, the control device 18 is in a state where the estimated slip angle βe and the actual slip angle βre are substantially equal within a certain error range, and the estimated yaw rate value γe is a positive value set in comparison with the actual yaw rate γre. If the value is larger than the threshold value, it is determined that the vehicle is in an understeer state, the understeer state is canceled, and a control signal is transmitted to various actuators 19 so as to prevent the occurrence of drift.

  In this case, for example, the control device 18 transmits a signal for adjusting the torque distribution of the rear wheels with respect to the front wheels to the torque distribution electronic control system to eliminate the understeer, and the vehicle traveling direction side of the four-wheel independent brakes. A signal is transmitted so as to increase the brake pressure to the brake corresponding to the front wheel of the vehicle to prevent the occurrence of drift or to eliminate the drift state.

  Further, the control device 18 is in a state where the estimated slip angle βe and the actual slip angle βre are substantially equal within a certain error range, and the estimated yaw rate value γe is a negative value set in comparison with the actual yaw rate γre. If it is smaller than the threshold value, it is determined that the vehicle is in an oversteer state, and, for example, a signal for adjusting the torque distribution of the front wheels with respect to the rear wheels to be increased is transmitted to the torque distribution electronic control system to eliminate the oversteer. .

  In addition to this, the vehicle stabilization control system 17 also provides, for example, predicted values γet + 1, γet + 2, γet + after 1 second, 2 seconds, and 3 seconds after the yaw rate from the vehicle behavior estimation prediction device 1. 3 is output, and the control device 18 monitors the predicted values γet + 1, γet + 2, γet + 3 and the current steering angle δt of the steering wheel.

  At this time, for example, in a state where the steering angle δt is a positive or negative value and the driver turns the steering wheel, the predicted yaw rate γet + 3 after 3 seconds becomes the predicted yaw rate γet + 1 after 1 second, If it is predicted to change rapidly from γet + 2, it is determined that spin will occur after 3 seconds.

  In addition, the brake pressure is increased in advance for the brake located outside the turning of the vehicle among the four-wheel independent brakes, the throttle opening is throttled for the electronically controlled throttle that controls the accelerator opening, and the steering servo It is also possible to configure the system having a function to prevent the occurrence of spin by transmitting a signal so as to return the steering angle of the steering angle to a smaller direction.

  Further, for example, the vehicle behavior estimation / prediction device 1 is configured to output an estimated value Gye of the vehicle body lateral acceleration and a predicted value Gyet + a, Gyet + b, Gyet + c,. It is also possible to make a configuration so as to compare this value with the actual measurement value Gyre from the lateral acceleration sensor S6.

  If comprised in this way, the control apparatus 18 will compare those values with the measured value Gyre by lateral acceleration sensor S6, and a lateral acceleration will be carried out to a vehicle by the strong side wind, the inclination of the left-right direction of a road surface, or a drift. It can be determined that it has occurred. In that case, a control signal is transmitted to the system having the steering servo function so as to cut the steering hole in such a direction that the difference between the measured value Gyre of the lateral acceleration and the estimated value or the predicted value Gye is reduced. Is possible.

  As described above, according to the vehicle behavior estimation / prediction device 1 according to the present embodiment, in the hourglass neural network inSNN of the feature quantity extraction unit 2, the handle angle δ and the like are input to the input layer, and once the dimensions are compressed by the compression layer. After that, the handle angle δ and the like are restored from the output layer and output. Therefore, it can be said that the output from the compression layer to the output side intermediate layer is information that reflects all the features of the quantities representing a plurality of vehicle behaviors such as the input steering wheel angle δ.

  Accordingly, by extracting the information from the compression layer, extracting it as a feature value, and inputting it to the calculation means 3, an amount representing a plurality of vehicle behaviors such as the steering wheel angle δ in the neural network of the calculation means 3 with one input. It is possible to obtain the same effect as inputting. Therefore, it is possible to obtain the same effect as inputting many other types of quantities that may affect fluctuations in estimated values and predicted values representing the vehicle behavior output by the neural network to the neural network. Thus, it is possible to improve the accuracy of the estimated value and the predicted value of the quantity representing the vehicle behavior output from the neural network.

  In addition, it is possible to reduce the number of data that is actually input to the neural network by inputting the number representing the vehicle behavior into a small number of feature quantities, and reducing the number of data that is actually input to the neural network. Can also be reduced. As a result, the number of variables to be optimized by the genetic algorithm, etc. is small, and it is easily converged in the learning of the neural network, the learning efficiency is improved and the learning speed is promoted, and the vehicle behavior output from the neural network is It is possible to further improve the accuracy of the estimated value and the predicted value of the amount to be expressed.

  Further, the calculation means 3 that calculates and outputs an estimated value or a predicted value of the amount representing the vehicle behavior is constituted by a recurrent neural network RNN having delay element feedback, and a plurality of current or past output points are output. The estimated value or predicted value is dimensionally compressed using an hourglass neural network and input to the input layer of the recurrent neural network RNN.

  For this reason, the estimated value or predicted value output from the calculation means 3 is reflected on the estimated value or predicted value output immediately before or several cycles before, and the output estimated value or predicted value has continuity or causality. It becomes possible to have sex. In addition, since it is possible to dimensionally compress information that reflects the output characteristics of the past several cycles output in the past into one or a small number of information and input it to the recurrent neural network RNN, The same effect as described above can be obtained.

  In learning, by individually learning the hourglass neural network inSNN, retSNN and recurrent neural network RNN while switching the learning timing, each learning easily converges, learning efficiency is improved, and learning speed is increased. In addition to being promoted, it is possible to further improve the accuracy of the estimated value and the feature amount representing the vehicle behavior output from the neural network.

  Further, the learning timing is switched based on the average value, variance, and period of the waveform of the estimated value and predicted value output data representing the vehicle behavior output from the computing means 3 and the teacher data used for learning. If the difference between the average values of the waveforms is large, the learning of the hourglass neural network inSNN is bad, and if the variance of the respective waveforms is large, the learning of the recurrent neural network RNN is bad, and the period of each waveform. If the difference is large, learning of the hourglass neural network retSNN is bad.

  Therefore, it is possible to appropriately select the neural network to be learned by switching the learning timing based on the average value, variance, and period of each waveform of the output data and the teacher data, and the vehicle behavior estimation / prediction device 1 It becomes possible to improve the accuracy of the quantity representing the vehicle behavior to be output efficiently and accurately.

  On the other hand, according to the vehicle stabilization system 17 according to the present embodiment, the current behavior of the vehicle is estimated based on the highly accurate estimated value or predicted value transmitted from the vehicle behavior estimation / prediction device 1, and the future behavior is estimated. Can be predicted. Therefore, it is possible to effectively control the factor that destabilizes the vehicle behavior using the highly accurate estimated value and predicted value, or to correct the factor that destabilizes the vehicle behavior and Thus, it is possible to effectively control in advance so as not to occur.

  In addition, the calculation means of the vehicle behavior estimation / prediction device preliminarily learns the driving data in the state where the road surface is dry and the disturbance is small as the teacher data, thereby estimating the amount representing the vehicle behavior in the low disturbance state By outputting values and predicted values and comparing them with actual measured values, it is possible to grasp the disturbances that may or may occur in the vehicle and accurately control the stabilization of vehicle behavior Become.

  In this embodiment, the case where the steering wheel angle δ and the like are input to the vehicle behavior estimation / prediction device 1 has been described. However, the input data is not limited to these, and is input by appropriately connecting a sensor or the like to the device 1. Is done. Further, the data for calculating the differential value or the second-order differential value is not limited to the handle angle δ as in the present embodiment, but is determined as appropriate.

  Further, the type of data input to the feature quantity extraction unit 2 is appropriately determined according to the function to be performed by the vehicle behavior estimation / prediction device 1. In addition, as data to be input to each hourglass neural network of the feature amount extraction means, usually, amounts representing vehicle behaviors that are considered to have a strong correlation with each other are combined, but the combination of data is appropriately determined.

  Further, in the present embodiment, the case where the vehicle behavior estimation / prediction device 1 and the vehicle stabilization control system 17 using the vehicle behavior estimation / prediction device 1 are mounted on an actual vehicle has been described. It is also possible to configure, and such a case is also included in the present invention.

It is a block diagram which shows the structure of the vehicle behavior estimation prediction apparatus which concerns on this embodiment. It is a figure which shows the structure of the feature-value extraction means which concerns on this embodiment, (A) shows the case where the node of the compression layer of an hourglass type | mold neural network is single, and (B) shows the case where it is plurality. It is a figure which shows the structure of each calculating means which concerns on this embodiment. It is a figure which shows the structure of the optimization means which concerns on this embodiment. It is a block diagram which shows the structure of the optimization means which concerns on this embodiment. It is a block diagram which shows the structure of the vehicle stabilization control system using the vehicle behavior estimation prediction apparatus which concerns on this embodiment.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Vehicle behavior estimation prediction apparatus 2 Feature-value extraction means 3 Calculation means 9 Switching means 17 Vehicle stabilization system 18 Control apparatus S Measuring apparatus RNN Recurrent type neural network
inSNN hourglass neural network
retSNN hourglass neural network L3 compression layer L6 input layer L12 compression layer f1, f2 feature quantity D actual measurement data T teacher data Q evaluation value M average value σ variance F period

Claims (5)

Feature quantity extraction means for dimensionally compressing data relating to quantities representing a plurality of vehicle behaviors measured by the measuring device using an hourglass neural network and extracting an output from the compression layer of the hourglass neural network as a feature quantity;
An arithmetic means configured by a neural network that outputs an estimated value of the amount representing the current vehicle behavior based on the extracted feature amount or a predicted value of the amount representing the vehicle behavior after a certain time , and
The neural network constituting the calculation means is a recurrent neural network having a delay element feedback,
The recurrent neural network having the delay element feedback includes the hourglass neural network for dimensionally compressing the delay elements that are the estimated values or the predicted values output at a plurality of current or past time points, and the hourglass vehicle behavior estimation and prediction and wherein that you have been configured to input the output from the compression layer type neural network input layer of the recurrent neural network. The hourglass neural network for extracting the feature value, the recurrent neural network, and the hourglass neural network for dimensionally compressing the plurality of delay elements are measured data measured by a measuring device and output data from the recurrent neural network. vehicle behavior estimating prediction apparatus of claim 1, wherein each advance based on the evaluation value calculated as the teacher data, characterized in that it is learned individually. In learning of the hourglass neural network that extracts the feature quantity, the recurrent neural network, and the hourglass neural network that dimensionally compresses the plurality of delay elements, any one of them is performed in order to perform learning individually while switching the learning timing. The vehicle behavior estimation / prediction device according to claim 2 , further comprising switching means for selecting one. The switching means calculates an average value, variance, and period for the actual measurement data measured by the measurement device and the output data from the recurrent neural network, respectively, and normalizes the difference between the average value, variance, and period, respectively. When the difference between the normalized average values is the largest, the hourglass neural network that extracts the feature amount is used, and when the normalized variance is the largest, the recurrent neural network is used. The vehicle behavior estimation / prediction device according to claim 3 , wherein an hourglass neural network that dimensionally compresses the plurality of delay elements is selected when the normalized period difference is the largest. The vehicle behavior estimation / prediction device according to any one of claims 1 to 4 ,
A vehicle stabilization control system comprising: a control device that performs vehicle stabilization control based on the estimated value or the predicted value output from the vehicle behavior estimation prediction device.

Images