JP6805977B2 - Vehicle control device - Google Patents

Description

The present invention relates to, for example, the technical field of a vehicle control device for controlling a vehicle in which a single operation pedal can be used as both a first type pedal and a second type pedal.

Vehicles usually have separate accelerator and brake pedals. An example of a vehicle control device that controls such a vehicle is described in Patent Document 1. Patent Document 1 discloses a vehicle control device that determines that a occupant mistakes the accelerator pedal for a brake pedal when the depression force of the accelerator pedal exceeds a predetermined threshold value.

Other prior art documents related to the present invention include Patent Documents 2 and 3.

JP 2015-000625 Japanese Unexamined Patent Publication No. 2012-144053 International Publication No. 2016/170786 Pamphlet

On the other hand, in the future, there is a possibility that a vehicle in which a single operation pedal can be used as both an accelerator pedal and a brake pedal will be developed. The vehicle control device that controls such a vehicle should determine whether the occupant is operating the operating pedal as an accelerator pedal or a brake pedal.

In this case, the vehicle control device can also make the above-mentioned determination by using an existing method for determining whether or not the passenger mistakes the accelerator pedal for the brake pedal. For example, the vehicle control device uses the method described in Patent Document 1 to determine whether the passenger is operating the operation pedal as an accelerator pedal or a brake pedal based on the depression force of the accelerator pedal. It is also possible to do. However, the operation content of the operation pedal may vary. For example, the operation content of the operation pedal may vary from passenger to passenger. Alternatively, for example, even if the operation contents of the operation pedals of the same passenger are different, there is a possibility that the operation contents may vary depending on the difference in the physical condition and mood of the passenger (or other factors). However, the vehicle control device described in Patent Document 1 described above does not consider such variations in the operation content of the operation pedal. Specifically, in the vehicle control device described in Patent Document 1 described above, the threshold value used for determining whether or not the passenger mistakes the accelerator pedal for the brake pedal is a variation in the operation content of the operation pedal. It is set without consideration. Therefore, simply using the method described in Patent Document 1 may deteriorate the accuracy of determining whether the passenger is operating the operation pedal as an accelerator pedal or a brake pedal. Problems arise.

It should be noted that such a technical problem is that not only a vehicle control device that controls a vehicle in which a single operation pedal can be used as an accelerator pedal and a brake pedal but also a single operation pedal can be used as two different types of pedals. The same occurs in a vehicle control device that controls various vehicles.

The above-mentioned problems can be given as an example of the problems to be solved by the present invention. An object of the present invention is to provide a vehicle control device capable of appropriately controlling a vehicle in which a single operation pedal can be used as both a first type pedal and a second type pedal.

One aspect of the vehicle control device of the present invention is an acquisition means for acquiring operation information regarding the operation content of the operation pedal that can be operated by the passenger to drive the vehicle, and the acquisition means acquired by the acquisition means. Based on the operation information, it is determined whether the passenger is operating the operation pedal as a first type pedal or as a second type pedal different from the first type pedal. When it is determined that the determination means and the passenger are operating the operation pedal as the first type of pedal, the vehicle is controlled by the first control mode based on the operation information. When it is determined that the passenger is operating the operation pedal as the second type of pedal, the vehicle is controlled by a second control different from the first control mode based on the operation information. The determination means includes a control means for controlling in an embodiment, and the determination means operates the operation content when the operation pedal is operated as the first type pedal and the operation pedal is operated as the second type pedal. Whether at least one of the operation contents of the case is learned and the passenger operates the operation pedal as the first type pedal or the second type based on the result of the learning in addition to the operation information. Determine if it is operating as a pedal.

FIG. 1 is a block diagram showing a configuration of a vehicle according to the present embodiment. FIG. 2A is a graph showing an accelerator operation pattern, and FIG. 2B is a graph showing a brake operation pattern. FIG. 3 is a block diagram showing a configuration of an operation determination ECU. FIG. 4 is a block diagram showing the configuration of the feature extractor. FIG. 5 is a block diagram showing the configuration of the classifier. 6 (a) and 6 (b) are graphs showing the feature vector calculated by the feature extractor in space with respect to the feature vector, respectively. FIG. 7 is a graph showing a hyperplane that classifies the feature vector calculated by the feature extractor into two classes in space with respect to the feature vector. FIG. 8 is a block diagram showing a configuration of the controller. 9 (a) and 9 (b) are graphs showing variations in accelerator operation patterns and brake operation patterns for each passenger or due to differences in the physical condition of the passengers, etc., using feature vector. FIG. 10 (c) is a graph schematically showing a state of learning of a classifier (that is, learning of a hyperplane) in consideration of variations in accelerator operation patterns and brake operation patterns.

Hereinafter, embodiments of the vehicle control device will be described. Hereinafter, the vehicle 1 to which one embodiment of the vehicle control device of the present invention is applied will be described.

(1) Configuration of Vehicle 1 First, the configuration of vehicle 1 of the present embodiment will be described with reference to FIG. FIG. 1 is a block diagram showing the configuration of the vehicle 1 of the present embodiment.

As shown in FIG. 1, the vehicle 1 includes a dual-purpose pedal 11 which is a specific example of an “operation pedal”, an operation determination ECU 12 which is a specific example of a “vehicle control device”, and an engine ECU (Electronic Control Unit) 13. The brake ECU 14, the engine 15, and the brake mechanism 16 are provided.

The dual-purpose pedal 11 is a single operation pedal that can also be used as an accelerator pedal and a brake pedal. When the occupant operates the combined pedal 11 as an accelerator pedal, the combined pedal 11 becomes an operation pedal that the occupant can operate in order to adjust the driving force of the vehicle 1. On the other hand, when the passenger operates the combined pedal 11 as a brake pedal, the combined pedal 11 can be operated by the passenger (that is, can be stepped on) in order to adjust the braking force of the vehicle 1. It becomes an operation pedal.

The operation determination ECU 12 performs an operation determination operation for determining whether the passenger is operating the dual-purpose pedal 11 as an accelerator pedal or a brake pedal. In the present embodiment, the operation determination ECU 12 performs the operation determination operation based on the operation signal Sp indicating the operation content of the dual-purpose pedal 11 on the premise that there is a difference between the accelerator operation pattern and the brake operation pattern. .. The accelerator operation pattern is a time-series pattern of the operation signal Sp obtained when a passenger who intends to operate the combined pedal 11 as an accelerator pedal operates the combined pedal 11. The brake operation pattern is a time-series pattern of the operation signal Sp obtained when a passenger who intends to operate the combined pedal 11 as a brake pedal operates the combined pedal 11.

Hereinafter, the difference between the accelerator operation pattern and the brake operation pattern will be described with reference to FIGS. 2 (a) and 2 (b). FIG. 2A is a graph showing an accelerator operation pattern. On the other hand, FIG. 2B is a graph showing a brake operation pattern. In addition, in FIG. 2A and FIG. 2B, the operation signal Sp indicates "stroke amount (that is, depression amount) of the dual-purpose pedal 11" which is an example of a parameter indicating the operation content of the dual-purpose pedal 11. There is. Therefore, the accelerator operation pattern corresponds to a change pattern along the time axis of the stroke amount of the dual-purpose pedal 11 operated as an accelerator pedal. Similarly, the brake operation pattern corresponds to a change pattern of the stroke amount of the dual-purpose pedal 11 operated as the brake pedal along the time axis.

As shown in FIGS. 2A and 2B, for example, the accelerator pedal is relatively gently depressed between the accelerator operation pattern and the brake operation pattern, while the brake pedal is relatively. There is a difference that it is stepped on suddenly. That is, between the accelerator operation pattern and the brake operation pattern, the stroke amount increases or decreases relatively slowly in the accelerator operation pattern, while the stroke amount increases or decreases relatively rapidly in the brake operation pattern. There is a difference. Further, for example, there is a difference between the accelerator operation pattern and the brake operation pattern that the accelerator pedal is relatively difficult to be depressed deeply, while the brake pedal is relatively easy to be depressed. That is, there is a difference between the accelerator operation pattern and the brake operation pattern that the maximum value of the stroke amount is relatively small in the accelerator operation pattern, while the maximum value of the stroke amount is relatively large in the brake operation pattern. .. Further, for example, there is a difference between the accelerator operation pattern and the brake operation pattern that the accelerator pedal is continuously depressed for a relatively long period of time, while the brake pedal is depressed for a relatively short period of time. That is, between the accelerator operation pattern and the brake operation pattern, the period in which the stroke amount is larger than zero is relatively long in the accelerator operation pattern, while the period in which the stroke amount is larger than zero is relative in the brake operation pattern. The difference is that it is short.

In the operation determination ECU 12, the actual operation pattern (that is, the time-series pattern of the operation content of the dual-purpose pedal 11) indicated by the operation signal Sp input to the operation determination ECU 12 is the accelerator operation pattern (that is, corresponds to the accelerator operation pattern). For example, the operation determination operation is performed by determining whether it is the same or similar) or the brake operation pattern. When the actual operation pattern is the accelerator operation pattern (that is, corresponding to the accelerator operation pattern), the operation determination ECU 12 determines that the passenger is operating the combined pedal 11 as the accelerator pedal. When the actual operation pattern is the brake operation pattern (that is, corresponding to the brake operation pattern), the operation determination ECU 12 determines that the passenger is operating the combined pedal 11 as the brake pedal.

Again, in FIG. 1, the engine ECU 13 controls the engine 15 so as to adjust the driving force of the vehicle 1 based on the operation signal Sp when the combined pedal 11 is operated as an accelerator pedal. The brake ECU 14 controls the brake mechanism 16 so as to adjust the braking force of the vehicle 1 based on the operation signal Sp when the combined pedal 11 is operated as a brake pedal. The engine 15 is a drive source that generates a driving force for the vehicle 1. The brake mechanism 16 is a braking device that generates a braking force of the vehicle 1 by using an electric pressure or the like.

In this embodiment, in particular, the operation determination ECU 12 performs a learning operation for learning the operation signal Sp so as to improve the determination accuracy by the operation determination operation. Hereinafter, the configuration and operation of the operation determination ECU 12 that performs the operation determination operation and the learning operation will be described in order.

(2) Operation determination ECU 12
(2-1) Configuration of Operation Judgment ECU 12 First, the configuration of the operation determination ECU 12 will be described with reference to FIG. FIG. 3 is a block diagram showing the configuration of the operation determination ECU 12.

As shown in FIG. 3, the operation determination ECU 12 is a specific example of the “acquisition means” as a processing block logically realized in the operation determination ECU 12 or a processing circuit physically realized in the operation determination ECU 12. It includes a controller 121, a neural network 122 which is a specific example of "determination means", a memory 123, and a controller 124 which is a specific example of "control means". The controller 121, the neural network 122, the memory 123, and the controller 124 can communicate with each other via the data bus 125.

The controller 121 controls the entire operation of the operation determination ECU 12. In particular, the controller 121 acquires the operation signal Sp from the dual-purpose pedal 11. The controller 121 outputs the acquired operation signal Sp to the neural network 122. The controller 121 may store the acquired operation signal Sp in the memory 123.

The neural network 122 performs an operation determination operation based on the input operation signal Sp. Specifically, the neural network 122 calculates the probability p (or the probability that the actual operation pattern is the accelerator operation pattern) that the actual operation pattern indicated by the operation signal Sp is the brake operation pattern. In other words, the neural network 122 calculates the similarity of the actual operation pattern to the brake operation pattern (or the similarity to the accelerator operation pattern). In order to calculate the probability p (in other words, the similarity, the same applies hereinafter), the neural network 122 is based on the feature extractor 126 that calculates the feature vector C indicating the features of the actual operation pattern and the feature vector C. It is provided with a classifier 127 for calculating the probability p. The details of the feature extractor 126 and the classifier 127 will be described together with the explanation of the operation of the operation determination ECU 12, and thus the description thereof will be omitted here.

Further, the neural network 122 learns the input operation signal Sp. In particular, the neural network 122 learns the input operation signal Sp so as to adjust the parameter group PG including a plurality of parameters defining the characteristics of the classifier 127 to improve the calculation accuracy of the probability p. The adjusted parameter group PG is stored in the memory 123.

The controller 124 controls the engine ECU 13 and the brake ECU 14 based on the probability p calculated by the neural network 122. Since the details of the operation of the controller 124 will be described together with the explanation of the operation of the operation determination ECU 12, the description thereof will be omitted here.

(2-2) Operation of operation determination ECU 12
(2-2-1) Initial learning of the neural network 122 First, when the operation determination ECU 12 is mounted on the vehicle 1 (in other words, before the vehicle 1 is shipped to the market), the initial learning of the neural network 122 is performed. Will be. Specifically, first, a plurality of accelerator operation patterns and brake operation patterns are collected by the controller 121. The accelerator operation pattern may be collected by the test driver actually operating the combined pedal 11 as an accelerator pedal. The accelerator operation pattern may be collected by the test driver actually operating the accelerator pedal in a vehicle equipped with the accelerator pedal. Accelerator operation patterns may be collected from a simulation model that simulates the movement of a passenger depressing the accelerator pedal. The brake operation pattern may be collected by the test driver actually operating the combined pedal 11 as a brake pedal. The brake operation pattern may be collected by the test driver actually operating the brake pedal in a vehicle equipped with the brake pedal. Brake operation patterns may be collected from a simulation model that simulates the movement of the occupant depressing the brake pedal.

The controller 121 associates the collected plurality of accelerator operation patterns and the plurality of brake operation patterns with correct answer data indicating whether each operation pattern is an accelerator operation pattern or a brake operation pattern. Specifically, for example, the controller 121 indicates that each of the plurality of accelerator operation patterns (x _a1 , x _a2 , ..., X _aN ) is the correct answer data (t) indicating that these are the accelerator operation patterns. _{a1, t} a2, ···, associate _{t aN).} Note that N indicates the total number of accelerator operation patterns collected, and x _an (where n is _an integer satisfying 1 ≦ n ≦ N) is the nth accelerator operation pattern (that is, time-series data of the stroke amount). are shown, t _an, shows correct data indicating that the accelerator operation pattern x _an, is an accelerator operation pattern. Further, the controller 121 indicates, for example, for each of the plurality of brake operation patterns (x _b1 , x _b2 , ..., X _bM ) that the correct answer data (t _b1 , t) indicates that these are the brake operation patterns. _b2 , ···, t _bM ) are associated. Note that M indicates the total number of collected brake operation patterns, and x _bm (where m is an integer satisfying 1 ≦ m ≦ M) is the mth brake operation pattern (that is, time-series data of the stroke amount). Indicates, and t _bm indicates the correct answer data indicating that the brake operation pattern x _bm is the brake operation pattern.

After that, the controller 121 uses a plurality of accelerator operation patterns (x _a1 , x _a2 , ..., X _aN ), correct answer data (ta ₁ , ta ₂ , ..., _taN ), and a plurality of brake operation patterns (x _aN ). A matrix (x, t) is generated in which _b1 , x _b2 , ..., X _bM ) and correct data (t _b1 , t _b2 , ..., T _bM ) are concatenated. Note that x is a combination of a plurality of accelerator operation patterns (x _a1 , x _a2 , ..., X _aN ) and a plurality of brake operation patterns (x _b1 , x _b2 , ..., X _bM ). It is a obtained matrix, and is represented by a matrix (x _a1 , x _a2 , ..., X _aN , x _b1 , x _b2 , ..., X _bM ). Similarly, t is the correct answer data _{(t a1, t a2, ···} , t aN) and the correct answer data _{(t b1, t b2, ···} , t bM) in and is obtained by being linked matrix _{Yes, (t a1, t a2,} ···, t aN, t b1, t b2, ···, t bM) is expressed in that matrix. Therefore, the matrix (x, t) is a matrix indicating whether each matrix component (that is, operation pattern) of the matrix x is an accelerator operation pattern or a brake operation pattern.

Then, under the control of the controller 121, the neural network 122 is trained using the matrix (x, t). Specifically, the matrix component x _a1 constituting the matrix x is input to the feature extractor 126. The feature extractor 126 calculates a feature vector C _a1 indicating a feature of the matrix component x _a1 . The feature extractor 126 calculates the feature vector C _a1 using an existing algorithm for extracting the feature. However, for convenience of explanation, in the present embodiment, the feature extractor 126 calculates the feature vector C _a1 by performing a convolution process on the matrix component x _a1 . That is, in the present embodiment, the neural network 122 is a convolutional neural network (CNN: Convolutional Neural Network).

In this case, the feature extractor 126 includes an input interface 1261, a convolution processing unit 1262, and a pooling processing unit 1263, as shown in FIG. The convolution process performed by the convolution processing unit 1262 may be the same as the existing convolution process, and the pooling process performed by the pooling process unit 1263 may be the same as the existing pooling process. Therefore, for the sake of simplification of the description, detailed description of the convolution process and the pooling process will be omitted, but the outline thereof will be briefly described below. The matrix component x _a1 (that is, the operation signal Sp corresponding to the accelerator operation pattern) is input to the input interface 1261. The input interface 1261 converts the matrix component x _a1 , which is a serial data array, into a parallel data array. The input interface 1261 inputs the matrix component x _a1 , which is a parallel data array, to the convolution processing unit 1262. The convolution processing unit 1262 performs a convolution process on the matrix component x _a1 using a convolution filter having desired filter characteristics. At this time, the convolution processing unit 1262 may perform a plurality of convolution processes using a plurality of convolution filters having different filter characteristics on the matrix component x _a1 , or convolution using a single convolution filter. Processing may be performed. The data obtained as a result of the convolution process (so-called feature map) is input to the pooling process unit 1263. The pooling processing unit 1263 performs pooling processing on the feature map. As a result, the pooling processing unit 1263 outputs a D (where D is an integer of 1 or more) -dimensional feature vector C _a1 indicating the features of the matrix component x _a1 .

In the example shown in FIG. 4, the feature extractor 126 includes only one processing unit including a convolution processing unit 1262 and a pooling processing unit 1263. However, the feature extractor 126 may include a plurality of processing units including a convolution processing unit 1262 and a pooling processing unit 1263. In this case, the feature vector C _a1 output by the pooling processing unit 1263 of a certain processing unit is input to the convolution processing unit 1262 of the processing unit of the next stage. That is, the convolution process and the pooling process are repeated.

The feature amount vector C _a1 calculated by the feature extractor 126 is input to the classifier 127. The classifier 127 classifies the operation pattern indicated by the matrix component x _a1 into one of two classes based on the feature vector C _a1 . That is, the classifier 127 determines which of the two classes the operation pattern indicated by the matrix component x _a1 belongs to. Therefore, the classifier 127 is a two-class classification type classifier. Here, as described above, the neural network 122 calculates the probability p that the actual operation pattern indicated by the input accelerator operation signal Spa is the brake operation pattern. Therefore, the classifier 127 is configured to calculate the probability _pa1 that the operation pattern indicated by the matrix component x _a1 is the brake operation pattern. That is, the classifier 127 classifies the operation pattern indicated by the matrix component x _a1 into either a class corresponding to the accelerator operation pattern or a class corresponding to the brake operation pattern by calculating the probability _pa1 .

In such a classifier 127, for example, as shown in FIG. 5, all D input values (d1, d2, d3, ..., DD) constituting the D-dimensional feature vector C are input. It includes a coupling layer 1271 and an output layer 1272 that outputs a probability p based on the output from the fully coupled layer 1271.

The operation of calculating the probability p described above is also performed on the other matrix components x _a2 , ..., X _aN , x _b1 , x _b2 , ..., And x _bM that constitute the matrix x. As a result, the probability _p a2 operation pattern indicated by the matrix elements _{x a2} is the braking operation pattern, ..., the probability _{p aN} operation pattern indicated by the matrix elements _{x aN} is braking operation pattern, the operation indicated by the matrix elements _{x b1} probability p _b1 pattern is braking operation _pattern, ..., the probability p _bM operation pattern indicated by the matrix elements x _bM is braking operation pattern is calculated.

Thereafter, the controller 121, the probability classifier 127 is calculated _{p (= (p a1, p} a2, ···, p aN, p b1, ···, p bM)) and the correct answer data _{t (=} (t _a1 , T _a2 , ..., t _aN , t _b1 , ..., t _bM ))), the error function E (p, t) indicating the error is calculated. After that, the controller 121 trains the feature extractor 126 and the classifier 127 so that the error function E (p, t) is minimized. That is, the controller 121 trains the feature extractor 126 and the classifier 127 using a supervised learning algorithm. In this case, since the feature extractor 126 and the classifier 127 constitute the neural network 122, the controller 121 may train the feature extractor 126 and the classifier 127 by using the error back propagation method. However, the controller 121 may be trained by the feature extractor 126 and the classifier 127 using other algorithms. It should be noted that, from the probability operation pattern is a brake operation pattern is a probability p, correct data _{t a1, t a2, ···,} each of _{t aN} indicates "0 (= 0%)", the correct answer data _{Each of} t _b1 , t _b2 , ..., And t _bM indicates "1 (= 100%)". Further, since an existing error function (in other words, a loss function or an objective function) may be used as the error function E (p, t), a detailed description of the error function E (p, t) will be omitted.

Learning the feature extractor 126 may include adjusting the filter characteristics of the convolution filter used by the feature extractor 126. The information indicating the adjusted filter characteristics is stored in the memory 123 as the parameters constituting the parameter group PG. However, the learning of the feature extractor 126 may include adjustment of arbitrary characteristics of the feature extractor 126 (particularly, characteristics related to the calculation of the feature amount vector C). The learning of the classifier 127 may include adjusting the weights of a plurality of nodes N constituting the fully connected layer 1271. The learning of the classifier 127 may include adjusting the activation function (so-called softmax function) used by the output layer 1272. Information indicating the adjusted weight of the node N and the activation function is stored in the memory 123 as a parameter constituting the parameter group PG. However, the learning of the classifier 127 may include adjusting any characteristic of the classifier 127 (particularly, a characteristic related to the calculation of the probability p).

As a result of learning the feature extractor 126 so that the error function E (p, t) is minimized, the feature extractor 126 has the feature amount vector C as shown in FIG. 6A. The feature amount vector C can be calculated so that the feature amount vector C corresponding to the accelerator operation pattern and the feature amount vector C corresponding to the brake operation pattern are clearly separated in the corresponding D-dimensional space. Conversely, learning of the feature extractor 126 is performed so that the feature vector C corresponding to the accelerator operation pattern and the feature vector C corresponding to the brake operation pattern can be clearly separated. Will be done. As a result, in the feature amount vector C calculated by the feature extractor 126 before learning, as shown in FIG. 6B, the feature amount vector C corresponding to the accelerator operation pattern and the feature amount vector C corresponding to the brake operation pattern The feature vector C can be calculated so that the feature vector C and the feature vector C corresponding to the brake operation pattern are clearly separated by learning even when the features overlap at least partially. Become.

Further, as a result of the classifier 127 learning so that the error function E (p, t) is minimized, the classifier 127 is a D-dimensional space corresponding to the feature vector C as shown in FIG. Within, it is possible to specify a hyperplane (or a hyperplane, the same applies hereinafter) that is a determination boundary for separating the feature vector C corresponding to the accelerator operation pattern and the feature vector C corresponding to the brake operation pattern. .. That is, in the learning of the classifier 127 (that is, the adjustment of the weights of the plurality of nodes N constituting the fully connected layer 1272 and the adjustment of the activation function used by the output layer 1272), the actual operation pattern is the brake operation pattern. It is equivalent to the adjustment of the hyperplane corresponding to the calculation standard when calculating a certain probability p. Therefore, it can be said that the parameter group PG stored in the memory 123 substantially includes the parameters defining such a hyperplane. In particular, since the feature amount vector C corresponding to the accelerator operation pattern and the feature amount vector C corresponding to the brake operation pattern are clearly separated by the learning of the feature extractor 126, the classifier 127 corresponds to the accelerator operation pattern. It is possible to specify a hyperplane defined at a position away from both the feature amount vector C and the feature amount vector C corresponding to the brake operation pattern. Therefore, the probability p output by the classifier 127 is likely to be a value close to 0 or a value close to 1 (in other words, it is unlikely to be a value close to 0.5).

(2-2-2) Operation determination operation by the operation determination ECU 12 Then , referring to FIG. 3 again, the period during which the vehicle 1 is actually traveling (that is, the passenger actually operates the dual-purpose pedal 11). The operation determination operation performed by the operation determination ECU 12 during the period) will be described.

In this case, the operation signal Sp is sequentially input to the feature extractor 126. After that, the feature extractor 126 calculates the feature amount vector C (that is, the feature amount vector C indicating the feature of the actual operation pattern) indicating the feature of the operation signal Sp. The calculation operation of the feature vector C in the operation determination operation is the same as the calculation operation of the feature vector C in the initial learning.

After that, the classifier 127 calculates the probability p that the actual operation pattern indicated by the input operation signal Sp is the brake operation pattern based on the feature amount vector C. The calculation operation of the probability p in the operation determination operation is the same as the calculation operation of the probability p in the initial learning.

When the passenger is operating the dual-purpose pedal 11 as an accelerator pedal, the actual operation pattern should be the accelerator operation pattern. In this case, the classifier 127 classifies the feature amount vector C calculated by the feature extractor 126 into a class corresponding to the accelerator operation pattern. That is, the classifier 127 calculates the probability p close to 0. Therefore, the classifier 127 determines that the passenger is actually operating the dual-purpose pedal 11 as an accelerator pedal by calculating the probability p that is close to 0. On the other hand, when the passenger operates the dual-purpose pedal 11 as a brake pedal, the actual operation pattern should be the brake operation pattern. In this case, the classifier 127 classifies the feature amount vector C calculated by the feature extractor 126 into a class corresponding to the brake operation pattern. That is, the classifier 127 calculates a probability p close to 1. Therefore, by calculating the probability p close to 1, the classifier 127 determines that the passenger is actually operating the dual-purpose pedal 11 as a brake pedal.

The probability p calculated by the classifier 127 (that is, the determination result of whether the passenger is actually operating the dual-purpose pedal 11 as an accelerator pedal or a brake pedal) is input to the controller 124. To.

When the probability p is close to 0 (eg, less than the first predetermined threshold (eg, a value of 0.5 or less, 0.3, 0.2, 0.1, etc.)), the controller 124 , While outputting the operation signal Sp to the engine ECU 13, the operation signal Sp is not output to the brake ECU 14. In this case, the engine ECU 13 controls the engine 15 so as to supply a driving force according to the operation signal Sp. On the other hand, the brake mechanism 16 does not apply a braking force to the vehicle 1 according to the operation content of the dual-purpose pedal 11.

On the other hand, the controller 124 has a probability p close to 1 (for example, a first predetermined threshold value (for example, a value of 0.5 or more, 0.7, 0.8, 0.9, etc.) or more). In this case, the controller 124 outputs the operation signal Sp to the brake ECU 14, but does not output the operation signal Sp to the engine ECU 13. In this case, the brake ECU 14 controls the brake mechanism 16 so as to apply a braking force corresponding to the operation signal Sp to the vehicle 1. On the other hand, the engine 15 does not supply the driving force according to the operation signal Sp.

An example of the controller 124 capable of outputting the operation signal Sp to either the engine ECU 13 or the brake ECU 14 according to the value of the probability p will be described with reference to FIG. As shown in FIG. 8, the controller 124 includes a binarization processing unit 1241 and a switch 1242. The binarization processing unit 1241 binarizes the probability p. For example, the binarization processing unit 1241 converts the probability p close to 0 (for example, equal to or less than the first predetermined threshold value described above) to 0, and is close to 1 (for example, equal to or greater than the second predetermined threshold value described above). Convert the probability p to 1. An operation signal Sp is input to the switch 1242 from the controller 121. The switch 1242 sets the output destination of the operation signal Sp input to the switch 1242 to either the engine ECU 13 or the brake ECU 14 based on the output of the binarization processing unit 1241. Specifically, when the output of the binarization processing unit 1241 is 0, the switch 1242 sets the output destination of the operation signal Sp to the engine ECU 13. On the other hand, when the output of the binarization processing unit 1241 is 1, the switch 1242 sets the output destination of the operation signal Sp to the brake ECU 14.

(2-2-3) Online Learning of Neural Network 122 In this embodiment, the controller is further operated during the period in which the vehicle 1 is actually traveling (that is, the period in which the passenger is actually operating the dual-purpose pedal 11). The neural network 122 is learned by the operation determination ECU 12 using the operation signal Sp input to 121. That is, online learning (in other words, real-time learning) of the neural network 122 is performed. Specifically, each time the operation signal Sp is input to the neural network 122, the neural network 122 is learned. However, the neural network 122 may be learned each time the operation signal Sp is stored in the memory 123.

Online learning of the neural network 122 includes learning of the classifier 127. In this case, the controller 121 may train the classifier 127 using a supervised learning algorithm. That is, the controller 121 may train the classifier 127 in the same learning mode as the initial learning. In this case, the correct answer data may be input by, for example, the passenger. Alternatively, the controller 121 may train the classifier 127 using an unsupervised learning algorithm. An example of an unsupervised learning algorithm is the EM (Expectation Maximization) method. The online learning of the neural network 122 may include, in addition to or in lieu of, the learning of the classifier 127, the learning of the feature extractor 126. In this case as well, the controller 121 may train the feature extractor 126 using a supervised learning algorithm and / or an unsupervised learning algorithm. However, since the processing load required for learning of the feature extractor 126 may be larger than the processing load required for learning of the classifier 127, the feature extractor has the purpose of reducing the processing load required for learning. 126 learning (ie, online learning) need not be performed.

Online learning of the neural network 122 may be performed in the operation determination ECU 12. Alternatively, the online learning of the neural network 122 may be performed by a data center outside the vehicle 1. In this case, the operation signal Sp is transmitted from the vehicle 1 to the data center via the network. A calculation model equivalent to the neural network 122 is constructed in the data center. The data center simulates the learning of the neural network 122 by using the operation signal Sp and the calculation model. The learning result by the data center is transmitted from the data center to the operation determination ECU 12 via the network. The operation determination ECU 12 adjusts the characteristics of the neural network 122 (that is, the parameter group PG) based on the learning result of the data center.

(3) Technical Effect According to the operation determination ECU 12 described above, learning of the neural network 122 using the accelerator operation pattern and the brake operation pattern is performed. That is, the operation determination ECU 12 uses the learning results of the accelerator operation pattern and the brake operation pattern to determine whether the actual operation pattern is the accelerator operation pattern or the brake operation pattern (that is, the passenger uses the dual-purpose pedal 11 as the accelerator pedal. Whether it is operated as a brake pedal or as a brake pedal) can be determined. Therefore, the determination accuracy by the operation determination operation is improved as compared with the operation determination ECU of the comparative example in which the accelerator operation pattern and the brake operation pattern are not learned.

In particular, in the present embodiment, the online learning of the neural network 122 is performed even while the vehicle 1 is traveling. That is, the neural network 122 is optimized according to the actual operation content of the dual-purpose pedal 11 by the passenger. Therefore, even if the operation content of the dual-purpose pedal 11 varies from passenger to passenger due to driving proficiency or the like, the operation determination ECU 12 can appropriately perform the operation determination operation. Alternatively, even if the operation content of the dual-purpose pedal 11 varies due to differences in the physical condition and mood of the passenger (or other factors), the operation determination ECU 12 can appropriately perform the operation determination operation.

Specifically, for example, FIG. 9A is a graph showing a feature amount vector C calculated when the first passenger operates the dual-purpose pedal 11. On the other hand, FIG. 9B is a graph showing a feature amount vector C calculated when a second passenger different from the first passenger operates the dual-purpose pedal 11. The hyperplane shown in FIGS. 9A and 9B is a hyperplane identified by the initial learning of the neural network 122. As shown in FIGS. 9A and 9B, the operation content of the combined pedal 11 may vary from passenger to passenger due to driving proficiency and the like. In this case, the hyperplane identified by the initial learning is clearly defined as the feature amount vector C corresponding to the accelerator operation pattern of the first passenger and the feature amount vector C corresponding to the brake operation pattern of the first passenger. It is separable. On the other hand, the hyperplane identified by the initial learning clearly defines the feature vector C corresponding to the accelerator operation pattern of the second passenger and the feature vector C corresponding to the brake operation pattern of the second passenger. It cannot be said that it is separable. In such a situation, when the online learning of the neural network 122 is performed while the second passenger is operating the vehicle 1, the accelerator operation pattern of the second passenger is as shown in FIG. 9 (c). A hyperplane optimized for the second passenger is defined, which can clearly separate the feature vector C corresponding to the above and the feature vector C corresponding to the brake operation pattern of the second passenger. As a result, even if the operation content of the dual-purpose pedal 11 varies from passenger to passenger due to driving proficiency or the like, the operation determination ECU 12 can appropriately perform the operation determination operation.

Alternatively, for example, FIG. 9A may be a graph showing the feature amount vector C calculated when a passenger in good physical condition operates the dual-purpose pedal 11. On the other hand, FIG. 9B may be a graph showing the feature amount vector C calculated when the same passenger who is not in good physical condition operates the dual-purpose pedal 11. That is, FIGS. 9 (a) and 9 (b) may indicate that the operation content of the dual-purpose pedal 11 may vary depending on the physical condition of the occupant and the like. In this case, the hyperplane identified by the initial learning is the feature vector C corresponding to the accelerator operation pattern of the passenger in good physical condition and the feature vector C corresponding to the brake operation pattern of the passenger in good physical condition. It is clearly separable. On the other hand, the hyperplane identified by the initial learning has a feature vector C corresponding to the accelerator operation pattern of the passenger who is not in good physical condition and a feature vector C corresponding to the brake operation pattern of the passenger who is not in good physical condition. It cannot be said that it is clearly separable. In such a situation, if the online learning of the neural network 122 is performed while the sick passenger is operating the vehicle 1, as shown in FIG. 9 (c), the accelerator of the sick passenger A hyperplane optimized for sick passengers is defined, which can clearly separate the feature vector C corresponding to the operation pattern and the feature vector C corresponding to the brake operation pattern of the sick passenger. Will be done. As a result, even if the operation content of the dual-purpose pedal 11 varies due to differences in the physical condition of the passengers, the operation determination ECU 12 can appropriately perform the operation determination operation.

In the above description, the neural network 122 is a convolutional neural network. However, the neural network 122 may be another type of neural network. For example, the neural network 122 may be a recurrent neural network (RNN: Recurrent Neural Network). For example, the neural network 122 may be an LSTM (Long Short Term Memory).

In the above description, the dual-purpose pedal 11 can also be used as an accelerator pedal and a brake pedal. However, the combined pedal 11 may be used as any two types of pedals. In this case, the operation determination ECU 12 determines whether the passenger is operating the dual-purpose pedal 11 as a first type pedal or a second type pedal different from the first type pedal. Perform the operation judgment operation for. When the combined pedal 11 is operated as the first type of pedal, the vehicle 1 is controlled in the first control mode based on the operation signal Sp. When the combined pedal 11 is operated as a second type pedal, the vehicle 1 is controlled in a second control mode different from the first control mode based on the operation signal Sp.

The present invention is not limited to the above-described embodiment, and can be appropriately modified within the scope of claims and within the scope not contrary to the gist or idea of the invention that can be read from the entire specification, and the vehicle control device accompanied by such modification. Is also included in the technical scope of the present invention.

1 Vehicle 11 Combined pedal 12 Operation judgment ECU
121 Controller 122 Neural Network 123 Memory 124 Controller 126 Feature Extractor 127 Classifier 13 Engine ECU
14 Brake ECU
15 engine 16 brake mechanism

Claims (1)

It is operable to the occupant to drive the vehicle, is not separately configured as a first type pedal and a second type pedal different from the first type pedal, and said An acquisition means for acquiring operation information regarding the operation contents of the passenger of a single operation pedal that can also be used as the first and second types of pedals , and
Based on the operation information acquired by the acquisition unit, determines determines whether the passenger has operated the operation pedal as the first type of or the second type are operating as a pedal of the pedal Means and
When it is determined that the passenger is operating the operation pedal as the first type of pedal, the vehicle is controlled by the first control mode based on the operation information, and the passenger When it is determined that the operation pedal is operated as the second type of pedal, the vehicle is controlled by a second control mode different from the first control mode based on the operation information. Equipped with control means
The determination means learns at least one of the operation content when the operation pedal is operated as the first type pedal and the operation content when the operation pedal is operated as the second type pedal. Based on the learning result in addition to the operation information, it is determined whether the passenger is operating the operation pedal as the first type pedal or the second type pedal. A vehicle control device characterized by the fact that.

Images