YAW RATE FORECASTING

Abstract

A yaw-rate forecasting system for a vehicle. The system includes a yaw rate sensor and an electronic control unit. The yaw rate sensor is configured to detect a yaw rate of the vehicle and to generate a signal indicative of the detected yaw rate. The electronic control unit is coupled to the yaw rate sensor, and is configured to receive the signal indicative of the yaw rate of the vehicle from the yaw rate sensor, forecast a future yaw rate, determine a stability of the vehicle using the forecasted yaw rate, and to generate a signal to control actuation of a vehicle brake.

Description

BACKGROUND

The invention generally relates to yaw rate forecasting. Specifically, the invention relates to use of Data Dependent Systems (DDS) Modeling, Autoregressive Moving Average models, and/or Autoregressive models, to forecast the yaw rate for use in Electronic Stability Control (ESC) systems.

ESC systems in vehicles use the yaw rate of the vehicle to make decisions regarding corrective actions. The yaw rate is provided to the ESC system by a sensor which detects the yaw rates and outputs a signal indicative of the detected yaw rate. In some systems, the yaw rate is provided to the ESC system over a bus (e.g., a CAN bus).

SUMMARY

Although current yaw rate sensors provide information to ESC systems, in most cases neither the sensors nor the ESC system forecast or predict what the yaw rate will be in the future. Forecasting yaw rate into the future can help account for communication and ESC building-pressure latencies. This can improve performance of the ESC system and reduce false interventions.

In one embodiment, the invention provides a yaw-rate forecasting system for a vehicle. The system includes a yaw rate sensor and an electronic control unit. The yaw rate sensor is configured to detect a yaw rate of the vehicle and to generate a signal indicative of the detected yaw rate. The electronic control unit is coupled to the yaw rate sensor, and is configured to receive the signal indicative of the yaw rate of the vehicle from the yaw rate sensor, forecast a future yaw rate, determine a stability of the vehicle using the forecasted yaw rate, and to generate a signal to control actuation of a vehicle brake.

In another embodiment the invention provides a method of stabilizing a vehicle based on a predicted future yaw rate of a vehicle. The method includes detecting a plurality of yaw rates of the vehicle over a period of time by a yaw rate sensor, providing an indication of the detected yaw rates to a controller, generating a model of a future yaw rate using the detected yaw rates, verifying an adequacy of the model, extending the model if the model is found to be inadequate, determining a predicted yaw rate for a moment in the future based on the model, providing the predicted yaw rate to an electronic stability control system, and determining, by the electronic stability control system, an action to take based at least in part on the predicted yaw rate.

Other aspects of the invention will become apparent by consideration of the detailed description and accompanying drawings.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a block diagram of a system for stabilizing a vehicle using a predicted yaw rate.

FIGS. 2A-2C are a diagram of a DDS modeling procedure.

DETAILED DESCRIPTION

Before any embodiments of the invention are explained in detail, it is to be understood that the invention is not limited in its application to the details of construction and the arrangement of components set forth in the following description or illustrated in the following drawings. The invention is capable of other embodiments and of being practiced or of being carried out in various ways.

FIG. 1 shows a block diagram of a system 100 for predicting (forecasting) future yaw rates and using the forecast yaw rates to stabilize a vehicle. The system 100 includes a yaw rate sensor (YRS) 105 and a controller such as an electronic control unit (ECU) 110. The ECU 110 includes a forecaster 115 and an electronic stability control (ESC) system 120. The forecaster 115, the ESC 120, or both, can be separate from or integrated into the ECU 110. The system 100, including subsystems such as the forecaster 115 and the ESC 120, can be implemented in hardware, software, or a combination of hardware and software. In some embodiments, one or more of the ECU 110, the forecaster 115, and the ESC 120 includes a processor (e.g., a microprocessor, microcontroller, ASIC, DSP, etc.) and memory (e.g., flash, ROM, RAM, EEPROM, etc.; i.e., a non-transitory computer readable medium), which can be internal to the processor, external to the processor, or both.

The forecaster 115 receives a signal from the yaw rate sensor 105. Using the signals received over a period of time, the forecaster 115 predicts or forecasts what the yaw rate will be at a point in the future. The forecaster 115 provides the predicted yaw rate to the ESC 120 which uses the predicted yaw rate to make decisions on when and how to intervene, and brake the wheels of the vehicle, to stabilize the vehicle.

The forecaster 115 uses modeling, specifically Data Dependent Systems (DDS) Modeling, Autoregressive Moving Average models, and/or Autoregressive models, to forecast the yaw rate at a point in the future.

DDS Models

DDS modeling is a statistical methodology for analyzing a time series of data. A time series is a set of data that is ordered in time. The forecaster 115 collects data (i.e., the yaw rate signal from the YRS 105) at uniform time intervals for the DDS modeling. DDS modeling relies on the fact that the order of the data is important and that current observations rely on past observations. This relationship is expressed as the autocorrelation between successive data points. The data can be broken down to the regression of two independent parts, one that depends on previous data and one that is an independent sequence. The simplest form of this regression (a first order autoregression) is

X _t =φ·X _t-1 ÷a _t  (1.1)

where,

• • X_t=Data at time t,
  • φ=model (autoregressive) parameter (constant),
  • X_t-1=Data at time t−1, and
  • a_t=sequence of uncorrelated variables.

The regression of Eq. 1.1 is, more particularly, an autoregressive model (e.g., an autoregression) and shows the dependence of the current value of X on its past values. The model is a first order autoregressive model or AR(1) model. A set of data may rely on more than just the previous observation and the autoregressive model can take a general form of AR(n)

X _t=φ₁·X_t-1÷φ₂·X_t-2÷ . . . ÷φ_n·X_t-n+a_t  (1.2)

The autoregressive model assumes that the a_t's are independent, that is a_t doesn't depend on a_t-1, a_t-2, etc. If the data violates this assumption, at can be decomposed using regression similarly to X_t. Take for example the AR(1) model, replacing at with a_t′

X _t=φ₁·X_t-1+a_t′  (1.3)

The a_t′ can be decomposed into three parts, its dependence on X_1-2, a_t-1, and an independent sequence a_t.

a _t′=φ₂·X_t-2−θ₁·a_t-1+a_t  (1.4)

where,

• • θ=average model (moving average) parameter (constant)

Substituting back into the AR(1) model:

X _t=φ₁·X_t-1÷φ₂·X_t-2÷−θ₁·a_t-1  (1.5)

This is a second order autoregressive and a first order moving average model or ARMA(2,1). The ARMA order can also be increased, so the general form of an ARMA(n,n−1) model becomes

X _t=φ₁·X_t-1+φ₂·X_t-2+ . . . +φ_n·X_t-n÷a_t−θ₁·a_t-1−θ₂·a_t-2− . . . −θ_n-1·a_t-1÷1  (1.6)

Model Parameters

The autoregressive model parameters (θ₁, θ₂, etc.) are obtained by using least squares estimates minimizing the sum of squares of a_t's. For the AR(1) example the parameter estimate becomes

$\begin{array}{cc}{\hat{\phi }}_1=\frac{\sum _{t=2}^NX_t·X_{t-1}}{\sum _{t=2}^NX_{t-1}^2}& \left(1.7\right)\end{array}$

where,

N=number of data points

For higher order AR models, an ordinary regression formula can be used for parameter estimation

{circumflex over (β)}=(X′·X)⁻¹·X′·Y  (1.8)

For the general AR(n) model, the parameter estimation becomes

$\begin{array}{cc}\underset{\_}{\hat{\phi }}={\left({\underset{\_}{X}}^{\prime }·\underset{\_}{X}\right)}^{-1}·{\underset{\_}{X}}^{\prime }·\underset{\_}{Y}\mathrm{where},\underset{\_}{\hat{\phi }}=\left[\begin{array}{c}{\hat{\phi }}_1\\ {\hat{\phi }}_2\\ ⋮\\ {\hat{\phi }}_n\end{array}\right]\underset{\_}{X}=\left[\begin{array}{cccc}{X}_n& X_{n-1}& \dots & X_1\\ X_{n+1}& X_n& \dots & X_2\\ ⋮& ⋮& & ⋮\\ X_{N-1}& X_{N-2}& \dots & X_{N-n}\end{array}\right]\underset{\_}{Y}=\left[\begin{array}{c}{X}_{n+1}\\ X_{n+2}\\ ⋮\\ X_N\end{array}\right]& \left(1.9\right)\end{array}$

Estimating the moving average model parameters (θ₁, θ₂, etc.) requires nonlinear least squares. This can be seen from an ARMA(2,1) example, where all of the observations up to t−1 are known. The ARMA (2,1) model

X _t=φ₁·X_t-1÷φ₂·X_t-2÷a_t−θ₁·a_t-1  (1.5)

To solve for X_t the value of a_t-1 must be known, so an expression for a_t-1 is obtained.

a _t-1 =X _t-1−φ₁·X_t-2−φ₂·X_t-3+θ₁·a_t-2  (1.10)

Substituting equation 1.10 into the ARMA model results in:

X _t=(φ₁−θ₁)X_t-1+(φ₂+θ₁·φ₁)X_t-2÷θ₁·φ₂·X_t-3−θ₁²·a_t-2+a_t  (1.11)

This is repeated for a_t-2 and then a_t-3, and so on. It can be seen that Eq. 1.11 is nonlinear, because the products and squares of φ₁, φ₂, and θ₁ are present. The nonlinear least squares method is used for parameter estimation to reduce to an iterative least square. The recursive process required for the ARMA parameter estimation is much longer, computationally and computer/processor resource intensive, than the AR parameter estimation.

Modeling Procedure

The DDS modeling procedure starts by fitting an AR(1) model. If this model isn't adequate, it can be concluded that the a_t's are not independent and that they depend on X_t-2, a_t-1 and a_t. This leads to an ARMA(2,1) model (a higher order model) as derived above. If the ARMA(2,1) proves to be inadequate, again the at's can modeled to be dependent on X_t-3, a_t-2, a_t-1 and a_t.

a _t′=φ₃·X_t-3−θ₁·a_t-1−θ₂·a_t-2+a_t  (1.12)

Substituting back into the inadequate ARMA(2,1) model:

X _t=φ₁·X_t-1÷φ₂·X_t-2÷φ₃·X_t-3÷a_t−θ₁·a_t-1−θ₂·a_t-2  (1.13)

The model is now an ARMA(3,2) model. If this proves to be inadequate the procedure can be repeated and a ARMA(4,3) created. This process shows how the ARMA(n,n−1) model is arrived at. Using an ARMA(n,n−1) model ensures that the dependence of a_t has been removed. The ARMA(n,n−1) model is:

X _t=φ₁·X_t-1÷φ₂·X_t-2÷ . . . ÷φ_n·X_t-n+a_t−θ₁·a_t-1−θ₂·a_t-2− . . . −θ_n-1·a_t-n÷1  (1.6)

FIGS. 2A-2C illustrates the DDS modeling procedure. Once an adequate ARMA model is found, the actual parameters can be examined for further model reduction. If one of the parameters (e.g., an autoregressive parameter or a moving average parameter) contains zero within its confidence interval, that parameter can be assumed to be zero and removed. For example, FIG. 2B shows that if θ₁ and θ₂ contain zero within their confidence interval the ARMA(3,2) model can be reduced to an AR(3) model. This type of reduction is very useful, because the moving average parameters require recursive calculation, while the autoregressive parameters can be calculated directly. Thus, for each level of the ARMA model, reduction to a lower level, or to an AR or MA model is attempted.

Model Adequacy

A model order is deemed adequate if a significant reduction in the residual sum of squares (RSS) is no longer seen with higher order models.

$\begin{array}{cc}RSS=\sum _{t=n+1}^Na_t^2& \left(1.14\right)\end{array}$

Two other tests for model adequacy are also employed, the F-Test and residual autocorrelations. The F-Test is a test used in linear regression. The F-Test tests if some of the model parameters are restricted to zero, and is used as a comparison between two models.

$\begin{array}{cc}F=\frac{A_1-A_0}{s}÷\frac{A_0}{N-r}\sim F\left(s,N-r\right)& \left(1.15\right)\end{array}$

where,

• • A₁=sum of squares of the unrestricted model,
  • A₁=sum of squares of the restricted model, and
  • F(s,N−r)=F-distribution with s and N−r degrees of freedom.

If the value of F obtained from the equation is greater than a predetermined threshold (e.g. the value for a 5% significance value obtained from an F-distribution table), then the higher model is significant. If the value is lower, then the smaller order model is significant.

The residual autocorrelations is a rough check on the independence of the a_t's. The following formula can be used to determine how small the autocorrelations should be.

$\begin{array}{cc}\sigma \left({\hat{\wp }}_k\right)\cong {\frac{1}{\sqrt{N}}\left[1+2\left({\hat{\wp }}_1^2+{\hat{\wp }}_2^2+\dots +{\hat{\wp }}_p^2\right)\right]}^{1/2},k>p& \left(1.16\right)\end{array}$

where,

σ({circumflex over (ρ)}_k)=standard error of estimated autocorrelations when ρ_k=0 for k>p

The {circumflex over (ρ)}_k's are asymptotically normally distributed, so they can be assumed to have a value of zero within 5% by verifying

|{circumflex over (ρ)}_k|<1.96·σ({circumflex over (ρ)}_k)  (1.17)

Another way to test the independence of the a_t's is to use the unified autocorrelations, {circumflex over (ρ)}_k/σ({circumflex over (ρ)}_k). When all of the unified autocorrelations are autocorrelations are within a predetermined band (e.g. ±1.96), the a_t's can be assumed to be independent. The output of the DDS program uses the absolute value of the unified autocorrelations, and indicative of the quantity of unified autocorrelations that are greater than two.

Forecasting

The DDS models can be used to forecast or predict future values using conditional expectation. Take for example the AR(1) model at t−1, X_t-1 is a known value:

X _t=φ₁X_t-1+a_t  (1.1)

Since the a_t is unknown the one step ahead forecast becomes:

{circumflex over (X)} _t-1(1)=φ₁·X_t-1  (1.18)

The error for the forecast is then:

e _t-1(1)=X_t−{circumflex over (X)}_t-1(1)=a_t  (1.19)

This equation shows that the forecasting errors are also the a_t's, so the least squares estimate of the model parameters that minimizes the sum of squares of a_t's also minimizes the forecasting error. The conditional expectation forecasts can also be extended to higher order models. For the general AR(n) model:

{circumflex over (X)} _t-1(1)=φ₁·X_t-1+θ₂·X_t-2+ . . . +φ_n·X_t-n  (1.20)

The DDS models are capable of providing arbitrary lag forecasts, so the user can choose the step ahead in time for their specific application.

Modeling

Following the modeling procedure outlined above, the forecaster 115 finds the best fit models for yaw rate data. The models are then used to forecast future yaw rates. The future yaw rates are provided to the ESC 120 and used for stability control.

The forecasted yaw rates can also be used by other functions that need yaw rate information. The models are adjusted to provide yaw rate forecasts that are indicative of yaw rates expected at a certain time in the future to compensate for latencies in the system receiving the predicted yaw rates, improving the performance of the system using the yaw rate.

Various features and advantages of the invention are set forth in the following claims.

Claims

1. A yaw-rate forecasting system for a vehicle, the system comprising: a yaw rate sensor configured to detect a yaw rate of the vehicle and generate a signal indicative of the detected yaw rate;an electronic control unit coupled to the yaw rate sensor, the electronic control unit configured to receive the signal indicative of the yaw rate of the vehicle from the yaw rate sensor, forecast a future yaw rate, determine a stability of the vehicle using the forecasted yaw rate, and generate a signal to control actuation of a vehicle brake.

2. The system of claim 1, the electronic control unit forecasts the future yaw rate by generating a model of a future yaw rate using the detected yaw rates;verifying an adequacy of the model;extending the model if the model is found to be inadequate;determining a predicted yaw rate for a moment in the future based on the model; anddetermining, an action to take based at least in part on the predicted yaw rate.

3. The system of claim 2, wherein the model takes the form Xt=φ1·Xt-1÷φ2·Xt-2÷ . . . ÷φn·Xt-n+at  (1.2)

4. The system of claim 2, wherein the model is a first order autoregressive model.

5. The system of claim 4, wherein the model is tested for adequacy and the model is increased one order of autoregression and one order of moving average when the model is not adequate.

6. The system of claim 5, wherein the testing for adequacy and the increasing the order of the autoregression and of the moving average is repeated until the model is found to be adequate.

7. The system of claim 6, wherein the parameters of the model found to be adequate are examined, and the model is reduced when one or more of the parameters is found to be zero.

8. The system of claim 7, wherein the order of the autoregression is reduced when an autoregression parameter is zero.

9. The system of claim 7, wherein the order of the moving average is reduced when a moving average parameter is zero.

10. The system of claim 7, wherein one or more autoregression parameters are obtained using least squares estimates.

11. The system of claim 7, wherein a higher order autoregression parameter is obtained using an ordinary regression formula.

12. The system of claim 7, wherein a moving average parameter is estimated using a nonlinear least squares.

13. The system of claim 2, wherein the model is determined to be adequate when a significant reduction in a residual sum of squares is no longer seen with a higher order model.

14. The system of claim 2, wherein the model is determined to be adequate when at least one of an F-test returns a significance value below a predetermined threshold, and a unified autocorrelation is within a predetermined band.

15. A method of stabilizing a vehicle based on a predicted future yaw rate of a vehicle, the method comprising: detecting a plurality of yaw rates of the vehicle over a period of time by a yaw rate sensor;providing an indication of the detected yaw rates to a controller;generating a model of a future yaw rate using the detected yaw rates;verifying an adequacy of the model;extending the model if the model is found to be inadequate;determining a predicted yaw rate for a moment in the future based on the model;providing the predicted yaw rate to an electronic stability control system; anddetermining, by the electronic stability control system, an action to take based at least in part on the predicted yaw rate.

16. The method of claim 15, wherein the model is a first order autoregression, and further comprising testing the model for adequacy, increasing one order of autoregression and one order of moving average when the model is not adequate.

17. The method of claim 15, further comprising examining the parameters of a model verified as adequate, and reducing the model when one or more of the parameters is found to be zero.

18. The method of claim 15, further comprising determining the model is adequate when a significant reduction in a residual sum of squares is no longer seen with a higher order model.

19. The method of claim 15, further comprising determining the model is adequate when at least one of an F-test returns a significance value below a predetermined threshold, a residual autocorrelation of variables is below a predetermined threshold, and a unified autocorrelation is within a predetermined band.