JP2019218018A - Road face state discrimination method and road face state discrimination device - Google Patents

Abstract

Provided is a road surface state determination method and a road surface state determination device that can ensure the road surface state determination accuracy even when the calculation amount of time expansion / contraction is reduced.
A time series waveform of tire vibration detected by an acceleration sensor is windowed at time T by windowing means, and a time series waveform of tire vibration for each time window is extracted to obtain a feature vector X _{i for} each time window. , _A kernel function K _A is calculated from the feature vector X _i for each time window and a reference feature vector Y _ASVJ that is a feature vector for each time window calculated for each road surface condition calculated in advance. At the time of calculation, the calculation time of the kernel function K _A is reduced by using Y _AKJ which is a reference feature vector whose Lagrange's undetermined multiplier λ is equal to or greater than a predetermined threshold value m as the reference feature vector _YAVJ. I did it.
[Selection diagram] FIG.

Description

  The present invention relates to a method and an apparatus for determining a road surface condition using only data of a time-series waveform of vibration of a running tire.

Conventionally, as a method of determining a road surface state using only data of a time series waveform of a tire vibration during traveling, a time window calculated from a time series waveform extracted by multiplying a time series waveform of a tire vibration by a window function is used. A method has been proposed in which a road surface state is determined using a kernel function calculated from a characteristic amount for each road surface and a reference characteristic amount, which is a characteristic amount for each time window, obtained in advance for each road surface state.
The reference feature amount is obtained by machine learning (SVM) using, as learning data, a feature amount for each time window calculated from a time series waveform of tire vibration previously obtained for each road surface condition (for example, see Patent Document 1). ).

JP 2014-35279 A

  However, although the time expansion and contraction is an operation necessary for comparing the acquired time-series waveforms, there is a problem that the calculation time is long and the processing becomes very heavy due to a large amount of calculation. .

  The present invention has been made in view of the conventional problems, and provides a road surface state determination method and a road surface state determination device that can secure the road surface state determination accuracy even when the amount of time expansion / contraction calculation is reduced. The purpose is to:

The present invention provides a step (a) of detecting a vibration of a tire during running, a step (b) of extracting a time-series waveform of the detected tire vibration, and a step of adding a predetermined time width to the time-series waveform of the tire vibration. (C) extracting a time-series waveform for each time window by multiplying by the window function, (d) calculating a feature amount from the time-series waveform for each time window, and calculating in the step (d). Calculating a kernel function from the calculated feature values for each time window and a reference feature value selected from the feature values for each time window calculated from the time series waveform of the tire vibration obtained in advance for each road surface condition ( e) and a step (f) of determining the state of the running road surface based on the value of the identification function using the kernel function, wherein the step (d) and the step (d) are performed. e) between (G) extracting, from the reference feature values, a reference feature value whose corresponding Lagrange undetermined multiplier is equal to or greater than a preset threshold value, as a calculation feature value; Is characterized in that a kernel function is calculated from the feature amount calculated in the step (d) and the calculation feature amount extracted in the step (g).
This makes it possible to reduce the number of reference features used for calculating the kernel function K (X, Y), thereby increasing the calculation speed while ensuring the road surface state determination accuracy.

Note that, as the feature vector X _i , the vibration level of a specific frequency band of the time-series waveform for each time window extracted by applying the window function, the time-varying variance of the vibration level of the specific frequency band, and the time Any one, a plurality, or all of the cepstrum coefficients of the sequence waveform may be used. Further, the vibration level of the specific frequency band may be a frequency spectrum of a time-series waveform for each time window extracted by applying the window function, or a time-series waveform for each time window extracted by applying the window function. The accuracy of determining the time-series waveform road surface state obtained through the filter can be improved.
Further, if the kernel function is a global alignment kernel function, a dynamic time warping kernel function, or an operation value of the kernel function, it is possible to improve the determination accuracy of the road surface state.

Further, the present invention provides a tire vibration detecting means disposed on an air chamber side of an inner liner portion of a tire tread portion for detecting vibration of a running tire, and the tire vibration detected by the tire vibration detecting means. Windowing means for windowing the time-series waveform of the predetermined time width to extract a time-series waveform of tire vibration for each time window, and a vibration level of a specific frequency in the extracted time-series waveform for each time window. And a feature amount calculating means for calculating a feature amount having a component of the vibration level as a component, and a time window calculated from a time series waveform of tire vibration for each road surface condition calculated in advance. Storage means for storing a reference feature quantity selected from the feature quantities of the above and a Lagrange undetermined multiplier corresponding to the reference feature quantity; a feature quantity for each time window calculated by the feature quantity calculation means; A kernel function calculating unit that calculates a kernel function from the stored reference feature amount; and a road surface state determining unit that determines a road surface state based on a value of an identification function using the kernel function. In the road surface state determination device for determining the state of the, from among the reference feature amounts stored in the storage means, a reference feature amount for which the corresponding Lagrange undetermined multiplier is equal to or greater than a preset threshold value, as a calculation feature amount In addition to providing calculation feature value extraction means for extraction, the kernel function calculation means calculates a kernel function from the feature value calculated by the feature value calculation means and the calculation feature value extracted by the calculation feature value extraction means. It is characterized in that it is calculated.
By adopting such a configuration, it is possible to obtain a road surface state determination device that can ensure the road surface state determination accuracy even when the amount of time expansion / contraction calculation is reduced.

  The summary of the invention does not list all the features required of the invention, and a sub-combination of these features may also be an invention.

FIG. 2 is a functional block diagram of the road surface condition determination device according to the first embodiment. It is a figure showing an example of a mounting position of an acceleration sensor. It is a figure showing an example of a time series waveform of tire vibration. It is a figure showing the method of calculating a feature vector from a time series waveform of tire vibration. It is a schematic diagram which shows an input space. It is a figure which shows the road surface feature vector of a DRY road surface and the road surface feature vector of a WET road surface in an input space. FIG. 4 is a diagram illustrating a method for calculating a GA kernel. 4 is a flowchart illustrating a road surface state determination method according to the present invention. It is a figure showing the distribution state of a support vector to a Lagrange multiplier. FIG. 6 is a diagram illustrating a relationship between a selection criterion of a feature vector for calculation and road surface determination accuracy.

Embodiment FIG. 1 is a diagram illustrating a configuration of a road surface state determination device 10 according to the present embodiment.
The road surface condition determination device 10 includes an acceleration sensor 11 as a tire vibration detection unit, a vibration waveform extraction unit 12, a windowing unit 13, a feature vector calculation unit 14, a storage unit 15, a kernel function calculation unit 16, The vehicle includes a road surface state determination unit 17 and a calculation feature amount extraction unit 18 as a data amount reduction unit, and performs two road surface determinations as to whether the road surface on which the tire 20 is traveling is a DRY road surface or a WET road surface. .
Each unit from the vibration waveform extracting unit 12 to the calculation feature amount extracting unit 18 is composed of, for example, computer software and a memory such as a RAM.
As shown in FIG. 2, the acceleration sensor 11 is disposed integrally at a substantially central portion of the inner liner portion 21 of the tire 20 on the tire air chamber 22 side, and detects vibration of the tire 20 due to input from a road surface. The tire vibration signal output from the acceleration sensor 11 is, for example, amplified by an amplifier, converted into a digital signal, and sent to the vibration waveform extracting means 12.

The vibration waveform extracting means 12 extracts a time series waveform of the tire vibration for each rotation of the tire from the signal of the tire vibration detected by the acceleration sensor 11.
FIG. 3 is a diagram showing an example of a time series waveform of the tire vibration. The time series waveform of the tire vibration has large peaks near the stepping position and the kicking position, and the land portion of the tire 20 is in contact with the ground. In the pre-stepping-in region R _f before the stepping-in, the kick-out region R _k after the land portion of the tire 20 is separated from the road surface, and the ground contact region R _s in which the land portion of the tire 20 is in contact with the road surface, Different vibrations appear depending on the state. On the other hand, the area before the stepping-in area _Rf and the area after the kicking-out area _Rk (hereinafter referred to as “out-of-road area”) are hardly affected by the road surface, so that the vibration level is small and the road surface information is low. Not included. Up area R _k after kicking from depression before area R _f, hereinafter referred to as the road surface area.

As shown in FIG. 4, the windowing means 13 windows the extracted time-series waveform with a predetermined time width (also referred to as a time window width) ΔT, and generates a time-series waveform of tire vibration for each time window. It is extracted and sent to the feature vector calculation means 14. Incidentally, T _s in the figure, a time width of the road area.
As described above, since the time-series waveform of the road surface area does not include the information of the road surface, in order to increase the calculation speed of the kernel function, in this example, only the time-series waveform of the road surface area is used as the feature vector calculation means. I send it to 14.
As the definition of the off-road area, for example, a background level may be set for a time-series waveform of tire vibration, and an area having a vibration level smaller than the background level may be set as the off-road area.

As shown in FIG. 4, the feature vector calculation means 14 calculates the feature vector X _i (i = 1 to N; N is the time for each extracted time window) Number of sequence waveforms).
In this example, as the feature vector X _{i to be} calculated, the time-series waveforms of the tire vibration are band-pass filters of 0-0.5 kHz, 0.5-1 kHz, 1-2 kHz, 2-3 kHz, 3-4 kHz, and 4-5 kHz, respectively. The vibration level (power value of the filtered wave) a _ik (k = 1 to 6) of the specific frequency band obtained by passing through each of them was used. Feature vectors, the number of _{X i = (a i1, a} i2, a i3, a i4, a i5, a i6) , the feature vector X _i is the N.
Figure 5 is a schematic diagram showing the input space of feature vectors X _i, each axis represents the vibration level a _ik of a specific frequency band, which is a feature quantity, each point representing a feature vector X _i. The actual input space is a seven-dimensional space when combined with the time axis because the number of specific frequency bands is three. However, the figure is expressed in two dimensions (the horizontal axis is a ₁ and the vertical axis is a ₂ ).
In the figure, a set of feature vectors X _i when the group C travels the DRY road, when group C be the set of _i 'is the feature vector X of when traveling the WET road', and the group C If the tire can be distinguished from the group C ′, it can be determined whether the road on which the tires are traveling is a DRY road surface or a WET road surface.

The storage means 15 stores a DW identification model for identifying a DRY road surface and a WET road surface, which has been obtained in advance.
The DW identification model includes a reference feature vector Y _ASV (y _jk ), which is a reference feature amount for separating a DRY road surface and a WET road surface by an identification function f (x) representing a separation hyperplane, and a reference feature vector Y _ASV ( y _jk ) and a Lagrange multiplier λ _A.
The reference feature values Y _ASV (y _jk ) and λ _A are the time series of tire vibration obtained by running a test vehicle equipped with a tire equipped with an acceleration sensor at various speeds on a DRY road surface and a WET road surface. The road surface feature vector Y _A (y _jk ), which is a feature vector for each time window calculated from the waveform, is obtained by learning using the input data as input data.
Note that the tire size used for learning may be one type or a plurality of types.
The subscript A of the reference feature vector Y _ASV (y _jk ) indicates DRY or WET.
The subscript j (j = 1 to M) indicates the window number of the time-series waveform extracted for each time window, and the subscript k indicates the vector component (k = 1 to 6). That is, _{y jk = (a j1, a} j2, a j3, a j4, a j5, a j6). SV is an abbreviation for support vector.
When a global alignment kernel function is used as in this example, the reference feature vector Y _ASV (y _jk ) is the number of dimensions of the vector y _i (here, 6 × M (M; number of windows)). Is a matrix.
Hereinafter, the road feature vector Y _A a (y _jk) and the reference feature vector Y _ASV (y _jk), respectively, referred to as Y _A, Y _ASV.

The method of calculating the road surface feature vector Y _A is the same as the feature vector X _j described above, for example, if the reference feature vectors Y _D of DRY road, the time-series waveform of tire vibrations when traveling along DRY road in time width ΔT and windowing, to extract the time-series waveform of tire vibrations per time window, it calculates a DRY road feature vector Y _D for each of the time-series waveform of the extracted each time window was. Similarly, the WET road surface feature vector Y _W is calculated from a time-series waveform for each time window when traveling on a WET road surface.
The reference feature vector Y _ASV is a feature vector selected as a support vector by a support vector machine (SVM) using the DRY road surface feature vector Y _D and the WET road surface feature vector Y _W as learning data.
Note that it is not necessary to store all the reference feature vectors Y _ASV in the storage means 15. Generally, only the support vector Y _ASV whose Lagrangian multiplier λ is equal to or more than a predetermined value λ _min (for example, λ _min = 0.05) is used. _May be stored as the reference feature vector Y _ASV .
Here, it is important that the time width ΔT has the same value as the time width ΔT when the feature vector _Xj is obtained. If the time width T is constant, the number M of time-series waveforms in the time window differs depending on the tire type and the vehicle speed. That is, the number M of the time-series waveforms in the time window of the road surface feature vector Y _ASV does not always match the number N of the time-series waveforms in the time window of the feature vector _Xj . For example, even if the tire type is the same, M> N when the vehicle speed at the time of finding the feature vector _Xj is lower than the vehicle speed at the time of finding the road surface feature vector Y _ASV, and M <N when the vehicle speed is fast. .

Figure 6 is a conceptual diagram showing a DRY road feature vector Y _D and WET road feature vector Y _W in the input space, black circles in the figure is DRY road, open circles are WET road.
As described above, although a DRY road feature vectors Y _D also WET road feature vector Y _W also matrices, for explaining how to determine the decision boundary of the group, in FIG. 6, DRY road feature vectors Y _D and WET The road surface feature vector Y _W is shown as a two-dimensional vector.
Group identification boundaries generally do not allow linear separation. Thus, by using the kernel method, the road surface feature vectors Y _V and Y _W are mapped to a high-dimensional feature space by a non-linear mapping φ to perform linear separation, so that the road surface feature vectors Y _D and Y _W are obtained in the original input space. Non-linear classification is performed.
In order to distinguish between a DRY road surface and a WET road surface, a margin is provided for an identification function f (x) that is a separating hyperplane that separates the DRY road surface feature vector Y _Dj and the WET road surface feature vector Y _Wj . The DRY road surface and the WET road surface can be accurately distinguished.
The margin refers to the distance from the separation hyperplane to the nearest sample, and the separation hyperplane that is the identification boundary is f (x) = 0. The DRY road surface feature vectors Y _Dj are all in the region of f (x) ≧ + 1, and the WET road surface feature vectors Y _Wj are all in the region of f (x) ≦ −1.

ここで、α，βは複数ある学習データの指標である。また、λはラグランジュ乗数で、λ＝０である路面特徴ベクトルＹ_Ajは、識別関数ｆ（ｘ）には関与しない（サポートベクトルではない）ベクトルデータである。
ここで、内積φ^T（ｘ_α）φ（ｘ_β）をカーネル関数Ｋ（ｘ_α，ｘ_β）に置き換えることで、識別関数ｆ（ｘ）＝ｗ^Tφ（ｘ）−ｂを非線形化できる。
なお、φ^T（ｘ_α）φ（ｘ_β）は、ｘ_αとｘ_βを写像φで高次元空間へ写像した後の内積である。
ラグランジュ乗数λは、前記の式（２）について、最急下降法やＳＭＯ（Sequential Minimal Optimization）などの最適化アルゴリズムを用いて求めることができる。
このように、内積φ^T（ｘ_α）φ（ｘ_β）を直接求めずに、カーネル関数Ｋ（ｘ_α，ｘ_β）に置き換えるようにすれば、高次元の内積を直接求める必要がないので、計算時間を大幅に縮減できる。 Next, using the data set X = (x ₁ , x ₂ ,..., X _n ) and the belonging class z = {1, −1}, the optimal identification function f (x) = w for identifying the data. Find ^T φ (x) -b. Here, w is a vector representing a weight coefficient, and b is a constant.
The data is DRY road feature vectors Y _Dj and WER road feature vector Y _Wj, with data DRY road indicated by chi ₁ belongs class z = 1 is the figure, WET indicated by z = -1 is chi ₂ This is road surface data. f (x) = 0 is the identification boundary, and 1 / || w || is the distance between the road surface feature vector Y _Aj (A = D, W) and f (x) = 0.
The discriminant function f (x) = w ^T φ (x) -b is optimized using, for example, the Lagrange multiplier method. The optimization problem is replaced by the following equations (1) and (2).

Here, α and β are indexes of a plurality of learning data. Further, λ is a Lagrange multiplier, and the road surface feature vector Y _Aj with λ = 0 is vector data that is not involved in the identification function f (x) (not a support vector).
Here, the discrimination function f (x) = w ^T φ (x) −b can be nonlinearized by replacing the inner product φ ^T (x _α ) φ (x _β ) with the kernel function K (x _α , x _β ). .
Note that φ ^T (x _α ) φ (x _β ) is an inner product after x _α and x _β are mapped to a high-dimensional space by a mapping φ.
The Lagrange multiplier λ can be obtained by using an optimization algorithm such as the steepest descent method or SMO (Sequential Minimal Optimization) for the above equation (2).
Thus, without asking the inner product phi ^T a _{(x α) φ (x β} ) directly, the kernel function K (x α, x _β) if so replace the, there is no need to determine directly the inner product of high dimensional , Can greatly reduce the calculation time.

ここで、||ｘ_αi−ｘ_βj||は、特徴ベクトル間の距離(ノルム)で、σは定数である。 In this example, a kernel function K (x _α, x _β) as was used global alignment kernel function (GA kernel).
As shown in FIG. 7 and the following equations (3) and (4), the GA kernel K (x _α , x _β ) is a local kernel κ _ij (indicating the similarity between the feature vector x _α and the feature vector x _β ) x _αi , x _βj ) can be directly compared with time series waveforms having different time lengths using a function composed of the sum or the sum of the products.
The local kernel κ _ij (x _αi , x _βj ) is obtained for each window of the time interval T.
FIG. 7 shows an example in which a GA kernel is obtained for a feature vector x _αi having six time windows and a feature vector x _β having four time windows.

Here, || x _αi −x _βj || is a distance (norm) between feature vectors, and σ is a constant.

本例では、選択基準１（λ≧０．３）とした。
これにより、データ量を削減できるので、カーネル関数Ｋ（Ｘ，Ｙ）の演算時間を速くすることができる。 The calculation feature amount extraction unit 18 is configured to select a reference feature for use in calculating a kernel function from the reference feature vector Y _DSV of the DRY road surface and the reference feature vector Y _WSV of the WET road surface recorded in the storage unit 15. The vectors Y _DK and Y _WK are selected and extracted, and sent to the kernel function calculating means 16 as a feature vector Y _DK for calculation of a DRY road surface and a feature vector Y _WK for calculation of a WET road surface, respectively.
As the selection criterion of the operation feature vector Y _AK , for example, there are the following three selection criterion.
Selection criterion 1: a reference feature vector Y _AK satisfying λ ≧ m is adopted.
However, m> λ _min = 0.05
Selection Criteria 2: Arrangement of λ is rearranged in descending order, and a fixed number N is adopted from the largest value.
Selection criterion 3; By rearranging the array of λ in descending order,
Calculate the ratio and evaluate what percentage of each λ occupies,
That is, the contribution rate is calculated and adopted.
Specifically, the contribution rate is added from the larger λ,
Up to λ where the integrated value of the contribution ratio exceeds k% (for example, 80%) is adopted.
Specific examples are shown in Table 1 below.

In this example, the selection criterion is 1 (λ ≧ 0.3).
As a result, the data amount can be reduced, and the calculation time of the kernel function K (X, Y) can be shortened.

The kernel function calculation means 16 calculates the feature vector X _i calculated by the feature vector calculation means 14, the DRY road calculation feature vector Y _DK extracted by the calculation feature quantity extraction means 18, and the WET road calculation feature. From the vector Y _WK , a DRYGA kernel K _D (X, Y _DK ) and a WETGA kernel K _W (X, Y _WK ) are calculated.
DRYGA kernel K _D (X, Y _DK), in the above formula (3) and (4), and a feature vector X _i calculated feature vector x _i.alpha the feature vector calculating section 14, DRY road feature vector x _beta local kernel _{κ ij (X i, Y DKj} ) when formed into a calculating feature vectors Y _DKj in total or function consisting of total product of, WETGA kernel K _W (X, Y _WK) is the feature vector x _beta WET This is a function composed of the sum or total product of the local kernels κ _ij (X _i , Y _Kj ) when the feature vector Y _WKj is used for calculating the road surface. By using the GA kernel K _D (X, Y _DK ) and the GA kernel K _W (X, Y _WK ), it is possible to directly compare time-series waveforms having different time lengths.
As described above, the number n of time-series waveforms in the time window when the feature vector X _i is obtained is different from the number m of time-series waveforms in the time window when the road surface feature vector Y _Aj is obtained. even if it is possible to determine the similarity between the calculating feature vectors Y _Wkj a feature vector X _i and the reference feature vector Y _ASVj.

ここで、Ｎ_DKはＤＲＹ路面の演算用特徴ベクトルＹ_DKjの個数で、Ｎ_WKはＷＥＴ路面の演算用特徴ベクトルＹ_WKjの個数である。
本例では、識別関数ｆ_DWを計算し、ｆ_DW＞０であれば、路面がＤＲＹ路面であると判別し、ｆ_DW＜０であれば、路面がＷＥＴ路面であると判別する。
演算用特徴ベクトルＹ_DKjの個数Ｎ_DK及び演算用特徴ベクトルＹ_WKjの個数Ｎ_WKは、それぞれ基準特徴ベクトルＹ_DSVの個数Ｎ_DSV及び基準特徴ベクトルＹ_WSVの個数Ｎ_WSVよりも少ないので、カーネル関数Ｋ（Ｘ，Ｙ）の演算時間を速くすることができる。
なお、後述する［実施例１］に示すように、演算用特徴ベクトルＹ_DKj及び演算用特徴ベクトルＹ_WKjの選定基準をλ≧０．３としても、ＤＲＹ/ＷＥＴの２路面の判別精度を十分確保することができる。 In the road surface state determining means 17, the value of the identification function f _DW (x) using the kernel function K _D (X, Y _DK ) and the kernel function K _W (X, Y _WK ) shown in the following equation (5) The road surface condition is determined based on the above.

Here, N _DK is the number of feature vectors Y _DKj for calculation on a DRY road surface, and N _WK is the number of feature vectors Y _WKj for calculation on a WET road surface.
In this example, the identification function f _DW is calculated, and if f _DW > 0, the road surface is determined to be a DRY road surface, and if f _DW <0, the road surface is determined to be a WET road surface.
Since the number N _WK number N _DK and calculating feature vectors Y _Wkj of calculation feature vector Y _DKj is less than the number N _WSV number N _DSV and the reference feature vectors Y _WSV each reference feature vectors Y _DSV, kernel function The calculation time of K (X, Y) can be shortened.
As described in [Example 1] described below, _{even if} the selection criterion of the calculation feature vector Y _DKj and the calculation feature vector Y _WKj is λ ≧ 0.3, the discrimination accuracy of the two road surfaces of DRY / WET is sufficient. Can be secured.

Next, a method of determining the state of the road surface on which the tire 20 is traveling using the road surface state determination device 10 will be described with reference to the flowchart of FIG.
First, a tire vibration generated by an input from a road surface R on which the tire 20 is traveling is detected by the acceleration sensor 11 (Step S10), and a time-series waveform of the tire vibration is extracted from the detected tire vibration signal (Step S10). S11).
Then, the time-series waveform of the extracted tire vibration is windowed over a predetermined time width ΔT to obtain a time-series waveform of the tire vibration for each time window. Here, the number of time-series waveforms of tire vibration for each time window is set to m (step S12).
Next, a feature vector X _i = (x _i1 , x _i2 , x _i3 , x _i4 , x _i5 , x _i6 ) is calculated for each of the extracted time-series waveforms in each time window (step S13). In this example, the time width T is 3 msec. The number of feature vectors X _i is six.
As described above, the components x _{i1 to} x _i6 (i = 1 to 6) of the feature vector X _i are the power values of the filter-filtered waves of the time series waveform of the tire vibration.
Next, a feature vector Y _DK for calculating the DRY road surface and a feature vector Y _DK for calculating the WET road surface are selected from the calculated feature vector X _i and the reference feature vector YA _SVj of the DRY road surface and the WET road surface recorded in the storage means 15. The feature vector Y _WK is extracted (step S14). Then, a local kernel κ _ij (X _i , Y _AKj ) is calculated from the calculation feature vector Y _DK and the calculation feature vector Y _WK and the feature vector X _i, and then the local kernel κ _ij (X _i , Y _AKj ) are calculated, and a GA kernel function K _A (X, Y _AK ) is calculated (step S15).
The GA kernel function K _D (X, Y _DK ) where A = D is a GA kernel function for a DRY road surface, and the GA kernel function K _W (X, Y _WK ) where A = W is a GA kernel function for a WET road surface. .
Then, an identification function f _DW (x) using the GA kernel function K _D of the DRY road surface and the GA kernel function K _W of the WET road surface is calculated (step S16). If f _DW > 0, the road surface is the DRY road surface And if f _DW <0, it is determined that the road surface is a WET road surface. (Step S17).
As described above, by adopting the reference feature vector Y _ADV that is λ ≧ m (> λ _min = 0.05) as the calculation feature vector Y _AK that is the reference feature vector used in the calculation of the kernel function, Since the data amount can be reduced while ensuring the determination accuracy, the road surface determination time can be effectively reduced.

図９（ａ）はサポートベクトルの分布を示す図で、図９（ｂ）はλ≧ｍであるサポートベクトルの個数を示す図である。図９（ｂ）に示すように、ｍを０．３とすれば、λ≧ｍ以上であるサポートベクトルの個数は、従来のｍ＝λ_min＝０．０５に比較して、約半分になることがわかる。
また、図１０は、ｍ＝０．０５，０．１，０．３，０．５，１．０としたときの、判別精度を示すグラフで、演算用特徴ベクトルＹ_AKの選択基準をλ≧０．３とした場合、判別精度は、従来のλ≧０．０５とした場合の９５％を確保しつつ、サポートベクトルの個数を４１５個から２６０個（およそ、４７％減）に減らすことができた。
また、同様に、計算時間を計測したところ、従来の４２％減を達成できた。
これにより、ラグランジェ未定乗数λが予め設定された閾値ｍ以上である基準特徴量を、カーネル関数を演算に使用する特徴量とすれば、路面状態の判別精度を確保しつつ、計算速度を速くすることができることが確認された。なお、従来の９０％以上の判別精度を確保するには、ｍ＜０．３５とすることが好ましい。 [Example]
A time window calculated from a time-series waveform of tire vibration when traveling on a DRY road surface and a WET road surface, in which the support vector of the DRY road surface and the support vector of the WET road surface have been obtained in advance on the DRY road surface and the WET road surface. The road surface data, which is the feature amount for each, was obtained as learning data by machine learning (SVM).
Specifically, as shown in Table 2 below, the used road surface data is divided into for training (for training) and for testing (for test), and the support vector for the DRY road surface and the support vector for the WET road surface are defined. After the determination, the support vector of the DRY road surface and the boundary surface of the support vector of the WET road surface were determined. At this time, the hyperparameters C and σ of the support vector machine were C = 2 and σ = 125, respectively.
At this time, the number of support vectors was 415 at the maximum.

FIG. 9A is a diagram illustrating the distribution of support vectors, and FIG. 9B is a diagram illustrating the number of support vectors satisfying λ ≧ m. As shown in FIG. 9 (b), when m is 0.3, the number of support vectors with λ ≧ m or more is about half as compared with the conventional m = λ _min = 0.05. You can see that.
FIG. 10 is a graph showing the discrimination accuracy when m = 0.05, 0.1, 0.3, 0.5, and 1.0. The selection criterion for the operation feature vector Y _AK is λ. When ≧ 0.3, the discrimination accuracy is to reduce the number of support vectors from 415 to 260 (approximately 47% reduction) while maintaining 95% of the conventional λ ≧ 0.05. Was completed.
Similarly, when the calculation time was measured, it was possible to achieve a 42% reduction compared to the conventional case.
As a result, if the reference feature amount in which the Lagrange's undetermined multiplier λ is equal to or larger than the preset threshold value m is used as a feature amount used in the calculation of the kernel function, the calculation speed is increased while ensuring the accuracy of determining the road surface state. It was confirmed that it could be done. Note that it is preferable to set m <0.35 in order to ensure the discrimination accuracy of 90% or more of the related art.

  As described above, the present invention has been described using the embodiment and the examples. However, the technical scope of the present invention is not limited to the scope described in the above embodiment. It will be apparent to those skilled in the art that various changes or improvements can be made to the embodiment. It is apparent from the appended claims that embodiments with such changes or improvements can be included in the technical scope of the present invention.

For example, in the above-described embodiment, two road surfaces are discriminated whether the road surface on which the tire 20 is traveling is a DRY road surface or a WET road surface using the DW identification model. Is used, it is possible to determine whether the road surface on which the tire 20 is traveling is a DRY road surface, a WET road surface, a SNOW road surface, or an ICE road surface.
Here, if A, A ′ = DRY, WET, SNOW, ICE (A ≠ A ′), the AA ′ identification model is an identification function f _{AA ′} (x ), A road surface feature vector Y _ASV and a Lagrangian multiplier λ _{AA ′} , which are reference feature amounts for separation by A), and an A ′ road surface feature vector Y _A′SV Lagrangian multiplier λ _A′A .
The reference feature values Y _ASV and λ _A are the values of tire vibration obtained by running a test vehicle equipped with a tire equipped with an acceleration sensor at various speeds on DRY, WET, SNOW, and ICE road surfaces. The road surface feature vector Y _A (y _jk ), which is a feature vector for each time window calculated from the series waveform, is obtained by learning using as input data.
The data of the A road surface, the data belonging to z = 1 indicated by chi ₁ in FIG. 6, data A 'road is data belonging to z = -1 shown by chi _2.

上記のように、識別関数がｆ_AA’（ｘ）であれば、Ａ路面のデータがｚ＝１に所属するデーで、Ａ’路面のデータがｚ＝−１に所属するデータであるので、６つの識別関数ｆ_AA’から、以下のように路面判別することができる。
ｆ_DW ＞０、ｆ_DS＞０、ｆ_DI＞０であれば、路面がＤＲＹ路面であると判別する。
ｆ_DW ＜０、ｆ_WS＞０、ｆ_WI＞０であれば、路面がＷＥＴ路面であると判別する。
ｆ_DS ＜０、ｆ_WS＞０、ｆ_SI＞０であれば、路面がＳＮＯＷ路面であると判別する。
ｆ_DI ＜０、ｆ_WI＜０、ｆ_SI＜０であれば、路面がＩＣＥ路面であると判別する。
なお、ＧＡカーネル関数Ｋ（Ｘ，Ｙ）に使用する特徴ベクトルを基準特徴ベクトルＹ_ASVとした場合には、式（６）〜（１１）において、Ｙ_AA’KをＹ_AA’SVとし、Ｎ_AA’KをＮ_AA’SVとすればよい。 It should be noted that there is a Lagrange multiplier λ _A corresponding to the reference feature vector Y _ASV for each identification model. For example, three Lagrangian multipliers λ _DW , λ _DS , and λ _DI corresponding to the DRY road surface feature vector Y _DSV have different values. The same applies to the other road surface feature vectors Y _WSV , Y _SSV and Y _ISV .
Here, the GA kernel function K _A (X, Y _AK ) is calculated using the feature vector used for the GA kernel function K (X, Y) as the calculation feature vector Y _AK .
The method of calculating the GA kernel function K _A (X, Y _AK ) is the same as in the first embodiment. The GA kernel function K _D (X, Y _DK ) in which A = D is the GA kernel function on the DRY road surface, and A = W GA kernel function K _W (X, Y _WK ) is a GA kernel function on a WET road surface, A = S GA kernel function K _S (X, Y _SK ) is a GA kernel function on a SNOW road surface, and A = I The GA kernel function K _I (X, Y _IK ) is the GA kernel function for the ICE road surface.
The determination of the road surface state is performed using the six identification functions f _{AA ′} (x) shown in the following equations (6) to (11).

As described above, if the discriminant function is f _{AA ′} (x), since the data on the road A is data belonging to z = 1 and the data on the road A ′ is data belonging to z = −1, From the six discriminant functions f _{AA ′} , the road surface can be determined as follows.
If f _DW > 0, f _DS > 0, f _DI > 0, it is determined that the road surface is a DRY road surface.
If f _DW <0, f _WS > 0, f _WI > 0, it is determined that the road surface is a WET road surface.
If f _DS <0, f _WS > 0, f _SI > 0, it is determined that the road surface is a SNOW road surface.
If f _DI <0, f _WI <0, f _SI <0, it is determined that the road surface is an ICE road surface.
When the feature vector used for the GA kernel function K (X, Y) is the reference feature vector Y _ASV , in equations (6) to (11), Y _AA′K is _set to Y _AA′SV and N _AA′SV is set. _AA'K may be _NAA'SV .

Further, in the above embodiment, the acceleration sensor 11 is used as the tire vibration detecting means, but other vibration detecting means such as a pressure sensor may be used. In addition, the acceleration sensor 11 may be installed at another location such as one at a position separated from the center of the tire in the width direction by a predetermined distance in the width direction, or may be installed in a block.
Further, in the embodiment, a feature vector X _i and the power value x _ik of filtration wave, variance, when the power value x _ik of filtration wave _{(log [x ik (t)} 2 + x ik ( t-1) ² ]) may be used. Alternatively, a feature vector X _i, Fourier coefficients a vibration level of a particular frequency band when the Fourier transform of the tire vibration time series waveform or may be cepstral coefficients. The cepstrum coefficient is obtained by assuming the waveform after Fourier transform as a spectrum waveform and performing Fourier transform again, or assuming the AR spectrum as a waveform and further obtaining the AR coefficient (LPC Cepstrum). Since the shape of the spectrum can be characterized without being affected by, the determination accuracy is improved as compared with the case where the frequency spectrum obtained by the Fourier transform is used.
In the above embodiment, the GA kernel is used as the kernel function, but a dynamic time warping kernel function (DTW kernel) may be used. Alternatively, a GA kernel and a DTW kernel operation value may be used.

10 road surface condition determination device, 11 acceleration sensor, 12 vibration waveform extraction means,
13 windowing means, 14 feature vector calculating means, 15 storage means,
16 kernel function calculating means, 17 road surface state determining means, 18 feature value extracting means for calculation,
20 tires, 21 inner liner part, 22 tire chamber.

Claims (4)

Detecting a vibration of the running tire (a), extracting a time-series waveform of the detected vibration of the tire (b), and applying a window function of a predetermined time width to the time-series waveform of the tire vibration. (C) extracting a time-series waveform for each time window by multiplying, (d) calculating a feature amount from the time-series waveform for each time window, and (E) calculating a kernel function from the characteristic amounts of the above and a reference characteristic amount selected from the characteristic amounts for each time window calculated from the time series waveform of the tire vibration obtained in advance for each road surface condition; (F) determining the state of the road surface on which the vehicle is traveling based on the value of the identification function using the kernel function;
In the road surface condition determination method provided with
Provided between the step (d) and the step (e),
(G) extracting, from the reference feature values, a reference feature value in which a corresponding Lagrange's undetermined multiplier is equal to or greater than a preset threshold value as a calculation feature value;
In the step (e),
A road surface state determination method, comprising calculating a kernel function from the feature amount calculated in the step (d) and the calculation feature amount extracted in the step (g). The feature quantity is
Vibration level of a specific frequency band of a time-series waveform for each time window extracted by applying the window function,
Time-varying dispersion of the vibration level of the specific frequency band,
And any one, or a plurality, or all of the cepstral coefficients of the time-series waveform;
The vibration level of the specific frequency band is a frequency spectrum of a time-series waveform for each time window extracted by applying the window function or a time-series waveform for each time window extracted by applying the window function through a band-pass filter. The road surface state determination method according to claim 1, wherein the vibration level is a vibration level in a specific frequency band obtained from the obtained time-series waveform.   3. The road surface state determination method according to claim 1, wherein the kernel function is a global alignment kernel function, a dynamic time warping kernel function, or an operation value of the kernel function. Tire vibration detection means arranged on the air chamber side of the inner liner portion of the tire tread portion to detect the vibration of the running tire,
Windowing means for extracting a time-series waveform of the tire vibration for each time window by windowing the time-series waveform of the tire vibration detected by the tire vibration detecting means with a preset time width,
A feature value calculating unit that calculates a feature value having a vibration level of a specific frequency as a component or a feature value having a function of the vibration level as a component in the extracted time-series waveform for each time window,
A storage for storing a reference feature value selected from feature values for each time window calculated from a time series waveform of tire vibration for each road surface condition calculated in advance and a Lagrange undetermined multiplier corresponding to the reference feature value. Means,
Kernel function calculation means for calculating a kernel function from the feature quantity for each time window calculated by the feature quantity calculation means and the reference feature quantity stored in the storage means;
A road surface state determination unit configured to determine a road surface state based on a value of a discrimination function using the kernel function, wherein a road surface state determination device configured to determine a state of a road surface on which a tire travels;
Calculating means for extracting, from among the reference features stored in the storage means, reference features whose corresponding Lagrange's undetermined multipliers are equal to or greater than a predetermined threshold value, as calculation features; ,
The kernel function calculating means,
A road surface state determination device, wherein a kernel function is calculated from the characteristic amount calculated by the characteristic amount calculation unit and the operation characteristic amount extracted by the operation characteristic amount extraction unit.

Images