7
$\begingroup$

I have a discrete function that represents a 1d position in space over time. I have a motion system that would like to achieve this curve, but, due to constraints on jerk (3rd derivative) and acceleration (2nd derivative), it cannot actually achieve the motion. This occurs due to a discontinuity in the function, where it goes from a flat signal to a sharp rise.

I'm looking for an algorithm that will take the discrete series and produce a new series that approximates the original, without violating the upper limits on the magnitude of the second and third derivatives. I don't have a strict definition for how I'd like to minimize the error for the position, but a standard sum mean-squared error fit is probably reasonable. For the derivative, I'm approximating using the standard centered difference $\frac{f(x+1)-f(x-1)}{(2{\Delta}t)}$.

I've tried simply applying the limits and driving toward the target position as fast as possible, but this is unstable (it doesn't properly consider the de-jerk and decel time and therefore overshoots the target and oscillates.) I've considered applying a repeated low-pass filter until the constraints are met, but this seems like a bit of a hack and I'm not convinced it will give me a very good fit. Is there some way to create a filter that specifically limits the derivatives of the signal?


A few additional notes:

  1. The new series must have the same length as the original series. This motion occurs at the same time as another operation and the two are strictly synchronized. It is acceptable, however, if it can't achieve the final position within the constraints; this just means that the parallel operation is too aggressive for the motion to follow. This should not occur, in practice, when configured properly.

  2. I've started investigating an iterative approach that identifies a point that violates the constraints, then adjusts for the error and splits the negative of the error between the two neighboring points, so that the sum of the function values will be preserved. It then computes the error at the neighbor and transfers the remaining error to the left or right, respectively. If it reaches the end of the function and there is still error, it just chops to the limits. I don't know if this approach has any sound theoretical grounding, but I'll experiment and see how it performs with my signals.


More notes:

It's reasonable to approximate the jerk (third derivative) function as a linear combination of non-overlapping step and impulse functions. For a typical input series, it looks something like this (the blue vs. red is my current attempt at filtering; this is mainly just to convey an idea of the shape):

Jerk Series

Here is a typical acceleration profile:

Accel Series

Here is a typical velocity profile (the climb at the start is what causes all the trouble; the physics of the problem make the system prefer to start moving at close to infinite velocity right from the get-go, but practical considerations get in the way):

Velocity Series

For those who are interested, here is the raw tab-delimited data series from which the central-difference derivatives were approximated to produce the above graphs:

x   4.5 4.5 4.5 4.5 4.5 4.5 4.5 4.5 4.5 4.5 4.5 4.5 4.5 4.5 4.5 4.5 4.5 4.5 4.5 4.5 4.5 4.5 4.5 4.5 4.5 4.5 4.5 4.5 4.5 4.5 4.5 4.5 4.5 4.5 4.510544858 4.529570168 4.548595479 4.567620789 4.5866461   4.60567141  4.624696721 4.643722031 4.662747342 4.681772652 4.700797963 4.719823273 4.738848584 4.757873894 4.776899204 4.795924515 4.814949825 4.833975136 4.853000446 4.872025757 4.891051067 4.910076378 4.929101688 4.948126999 4.967152309 4.98617762  5.00520293  5.024228241 5.043253551 5.062278862 5.081304172 5.100329483 5.119354793 5.138380104 5.157405414 5.176430725 5.195456035 5.214481346 5.233506656 5.252531967 5.271557277 5.290582588 5.309607898 5.328633209 5.347658519 5.36668383  5.38570914  5.404734451 5.423759761 5.442785072 5.461810382 5.480714657 5.497626044 5.514537431 5.531448818 5.548360206 5.565271593 5.58218298  5.599094367 5.616005754 5.632917141 5.649828528 5.666739915 5.683651302 5.700562689 5.717474077 5.734385464 5.751296851 5.768208238 5.785119625 5.802031012 5.818942399 5.835853786 5.852765173 5.869676561 5.886587948 5.903499335 5.920410722 5.937322109 5.954233496 5.971144883 5.98805627  6.004967657 6.021879044 6.038790432 6.055701819 6.072613206 6.089524593 6.10643598  6.123347367 6.140258754 6.157170141 6.174081528 6.190992916 6.207904303 6.22481569  6.241727077 6.258638464 6.275549851 6.292461238 6.309372625 6.326284012 6.343195399 6.360106787 6.377018174 6.393929561 6.410840948 6.427752335 6.444663722 6.461575109 6.478486496 6.495397883 6.512309271 6.529220658 6.546132045 6.563043432 6.579954819 6.596866206 6.613777593 6.63068898  6.647600367 6.664511754 6.681423142 6.698334529 6.715245916 6.732157303 6.74906869  6.765980077 6.782891464 6.79783353  6.807698506 6.817563482 6.827428457 6.837293433 6.847158409 6.857023385 6.866888361 6.876753336 6.886618312 6.896483288 6.906348264 6.91621324  6.926078215 6.935943191 6.945808167 6.955673143 6.965538119 6.975403095 6.98526807  6.995133046 7.004998022 7.014862998 7.024727974 7.034592949 7.044457925 7.054322901 7.064187877 7.074052853 7.083917828 7.093782804 7.10364778  7.113512756 7.123377732 7.133242707 7.143107683 7.152972659 7.162837635 7.172702611 7.182567587 7.192432562 7.202297538 7.212162514 7.22202749  7.231892466 7.241757441 7.251622417 7.261487393 7.271352369 7.281217345 7.29108232  7.300947296 7.310812272 7.320677248 7.330542224 7.3404072   7.350272175 7.360137151 7.370002127 7.379867103 7.389732079 7.399597054 7.40946203  7.419327006 7.429191982 7.439056958 7.448921933 7.45668417  7.46311785  7.46955153  7.47598521  7.48241889  7.488852569 7.495286249 7.501719929 7.508153609 7.514587289 7.521020969 7.527454649 7.533888329 7.540322008 7.546755688 7.553189368 7.559623048 7.566056728 7.572490408 7.578924088 7.585357768 7.591791447 7.598225127 7.604658807 7.611092487 7.617471594 7.622952136 7.628432678 7.63391322  7.639393762 7.644874304 7.650354846 7.655835389 7.661315931 7.666796473 7.672277015 7.677757557 7.683238099 7.688718641 7.694199183 7.699679725 7.705160268 7.71064081  7.716121352 7.721601894 7.727082436 7.732562978 7.73804352  7.743524062 7.749004605 7.754485147 7.759965689 7.765446231 7.770926773 7.776407315 7.781828236 7.786267475 7.790706714 7.795145953 7.799585192 7.804024432 7.808463671 7.81290291  7.817342149 7.821781388 7.826220627 7.830659866 7.835099105 7.839538344 7.843977584 7.848416823 7.852856062 7.857295301 7.86173454  7.866173779 7.870613018 7.875052257 7.879491497 7.883930736 7.888369975 7.892809214 7.897248453 7.901687692 7.906126931 7.91056617  7.915005409 7.919444649 7.923883888 7.928323127 7.932762366 7.937201605 7.941640844 7.946080083 7.950519322 7.954958561 7.959397801 7.96383704  7.968276279 7.972715518 7.977154757 7.980384921 7.982851165 7.985317409 7.987783653 7.990249897 7.992716141 7.995182385 7.997648629 8.000114873 8.002581117 8.005047361 8.007513605 8.009979849 8.012446092 8.014912336 8.01737858  8.019844824 8.022311068 8.024777312 8.027243556 8.0297098   8.032176044 8.034642288 8.037108532 8.039574776 8.04204102  8.044507264 8.046973508 8.049439752 8.051905996 8.05437224  8.056838484 8.059304728 8.061633119 8.063606114 8.065579109 8.067552105 8.0695251   8.071498095 8.07347109  8.075444085 8.07741708  8.079390076 8.081363071 8.083336066 8.085309061 8.087282056 8.089255051 8.091228046 8.093201042 8.09410447  8.094332124 8.094559777 8.09478743  8.095015084 8.095242737 8.09547039  8.095698043 8.095925697 8.09615335  8.096381003 8.096608657 8.09683631  8.097063963 8.097291616 8.09751927  8.097746923 8.097974576 8.09820223  8.098429883 8.098657536 8.098885189 8.099112843 8.099340496 8.099568149 8.099795803 8.100023456 8.100251109 8.100478762 8.100706416 8.100934069 8.101161722 8.101389376 8.101617029 8.101844682 8.102072336 8.102299989 8.102527642 8.102755295 8.102982949 8.103210602 8.103438255 8.103665909 8.103893562 8.104121215 8.104348868 8.104576522 8.104804175 8.105031828 8.105259482 8.105487135 8.105714788 8.105942441 8.106170095 8.106397748 8.106625401 8.106853055 8.107080708 8.107308361 8.107536014 8.107763668 8.107991321 8.108218974 8.108446628 8.108674183 8.108899792 8.109120089 8.109331659 8.109531087 8.109714959 8.10987986  8.110022376 8.110139158 8.11022965  8.110297009 8.110344649 8.110375986 8.110394434 8.110403409 8.110406324 8.110406324
$\endgroup$
  • $\begingroup$ Welcome to DSP.SE. Interesting question. $\endgroup$ – Jason R May 15 '13 at 17:04
  • $\begingroup$ How about some variant of a PID controller? $\endgroup$ – Jim Clay May 15 '13 at 18:00
  • $\begingroup$ @Jim, that might work, but it seems a bit complicated for the task and would require tuning to make it perform well. I'm also not sure it would be much better than a simple repeated low-pass filter. $\endgroup$ – Dan Bryant May 15 '13 at 18:09
  • 1
    $\begingroup$ I'm wondering whether taking the system described in this question/answer and, in the Kalman filter loop, forcing the jerk and acceleration to be constrained as you want might do the right thing. I have no time to play now; it'll be the weekend before I have a chance... the model state will need to be increased to include the jerk, as well as the constraints. $\endgroup$ – Peter K. May 16 '13 at 7:45
1
$\begingroup$

This is just an idea. Don't know if it will work. In cubic spline interpolation a curve is constructed by piece-wise cubic polynomials

$P(x) = a + bx + cx^2 + dx^3$

The coefficients for the $j^{th}$ cubic (which extends from the $j^{th}$ to the $(j+1)^{th}$ point) are found by directly specifying the first derivatives at the nodes and also specifying that the curve must pass through the points. I'm no where near an expert in this but it goes something like this

$P_j(x_j) = P_{j-1}(x_j) \\ P_j(x_{j+1}) = P_{j+1}(x_{j+1}) \\ P'_j(x_j) = P'_{j-1}(x_j) \\ P'_j(x_{j+1}) = P'_{j+1}(x_{j+1})$

Now using this idea to your problem maybe you could use $5^{th}$ order polynomials on this form

$P(x) = x^2(a + bx + cx^2 + dx^3)$

and use the constraints on the second and third order derivatives like this

$P''_j(x_j) = P''_{j-1}(x_j) \\ P''_j(x_{j+1}) = P''_{j+1}(x_{j+1}) \\ P'''_j(x_j) = P'''_{j-1}(x_j) \\ P'''_j(x_{j+1}) = P'''_{j+1}(x_{j+1})$

Edit: From the discrete function that represents the 1d position in space over time the acc and jerk values are computed (is that feasible in your case?). If they exceed some max value they are saturated to that max value. Then you construct a continuous 1D position function that has the acc and jerk values at the times you specified. Then this curve is evaluated at the times of interest and thereby you have a new discrete function that represents the 1d position in space over time but with the desired acc and jerk values. I don't know if this makes more sense. It takes some effort to test if it works. If you want to give it a try I suggest you search for cubic spline interpolation to learn how it works with position and velocity. Then use the same principles but just on acc and jerk. If I have to be more specific I will have to read up on cubic spline interpolation myself.

$\endgroup$
  • $\begingroup$ Is the intention of this to compute a piecewise continuous function that fits the discrete series? If so, do you have an idea of how to perform the jerk/accel limiting operation on the polynomials? If so, I can re-sample from the continous function, but it's unclear to me how I would modify the polynomial coefficients to achieve the desired maximal limit constraints. $\endgroup$ – Dan Bryant May 21 '13 at 17:16

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.