US-12617103-B2 - Systems and methods for task-specific adaptive velocity estimation in automated control and monitoring systems

Abstract

A method for adapting a causal FIR filter is provided. The causal FIR filter is configured to estimate velocities by filtering position measurement data in accordance with a frequency response of an electro-mechanical system involved in a task. The method uses a processor coupled with stored instructions to implement the method. The stored instructions, when executed by the processor, carry out steps of the method. The method includes collecting a sequence of position measurement data and a corresponding sequence of ground truth velocities of a component of the electro-mechanical system performing the task. The method further includes adapting the causal FIR filter for estimating velocities of the component performing the task by updating one or more coefficients of the causal FIR filter to reduce a difference between the sequence of ground truth velocities and estimated velocities produced by filtering the sequence of position measurement data by the causal FIR filter.

Inventors

• Daniel N. Nikovski
• William S. Yerazunis

Assignees

• MITSUBISHI ELECTRIC RESEARCH LABORATORIES, INC.

Dates

Publication Date

20260505

Application Date

20240206

Claims (18)

1. 1 . A method for adapting a causal finite impulse response (FIR) filter, wherein the causal FIR filter is configured to estimate velocities by filtering position measurement data in accordance with a frequency response of an electro-mechanical system involved in a task, wherein the method uses a processor coupled with stored instructions implementing the method, wherein the stored instructions, when executed by the processor, carry out steps of the method, comprising: collecting a sequence of position measurement data and a corresponding sequence of ground truth velocities of a component of the electro-mechanical system performing the task; adapting the causal FIR filter for estimating velocities of the component performing the task by updating one or more coefficients of the causal FIR filter to reduce a difference between the sequence of ground truth velocities and estimated velocities produced by filtering the sequence of position measurement data by the adapted causal FIR filter; and actuating the component of the electro-mechanical system according to a control command, wherein the control command is derived at least partially based on the estimated velocities.
2. 2 . The method of claim 1 , further comprising determining the sequence of ground truth velocities corresponding to the sequence of position measurement data by: processing the sequence of position measurement data with an acausal filter to generate a sequence of ground truth positions; and acausal central differencing the sequence of ground truth positions to obtain the sequence of ground truth velocities.
3. 3 . The method of claim 2 , wherein the acausal filter is a Savitzky-Golay symmetric acausal filter.
4. 4 . The method of claim 1 , further comprising determining the sequence of ground truth velocities corresponding to the sequence of position measurement data based on a sequence of ground truth positions, wherein the sequence of ground truth positions are determined by: solving an optimization problem to minimize a sum of a mean squared fitting error between the sequence of ground truth positions and the sequence of position measurement data, wherein the optimization problem is subject to box constraints determined by a quantization scheme of an encoder capturing the sequence of position measurement data and a term expressing a curvature of the sequence of ground truth positions; wherein the sequence of ground truth velocities are obtained by acausal central differencing the sequence of ground truth positions.
5. 5 . The method of claim 1 , wherein the ground truth velocities are measured via a tachometer.
6. 6 . The method of claim 1 , wherein the processor is configured to adapt coefficients of a first causal FIR filter for a first task performed by the component of the electro-mechanical system by comparing estimated velocities and ground truth velocities corresponding to the first task, and adapt coefficients of a second causal FIR filter for a second task performed by the component of the electro-mechanical system by comparing estimated velocities and ground truth velocities corresponding to the second task, wherein the first task and the second task are different from each other when at least one of a motion trajectory of the component or a load moved by the component are different between the first task and the second task.
7. 7 . The method of claim 1 , further comprising: receiving, via the electro-mechanical system, a second task to be performed by the component; collecting a second sequence of position measurement data and a second sequence of corresponding ground truth velocities corresponding to the second task; and updating the coefficients of the causal FIR filter to reduce a difference between the second sequence of corresponding ground truth velocities and the estimated velocities produced by filtering the second sequence of position data by the adapted causal FIR filter.
8. 8 . The method of claim 1 , wherein the coefficients of the causal FIR filter are updated using one or a combination of a linear regression fitting, a neural network execution, a moving least squares estimation, a radial basis function smoothing, a locally weighted scatterplot smoothing (LOWESS), a smoothing spline learning, and a Tikhonov smoothing.
9. 9 . The method of claim 1 , wherein the task corresponds to control of or estimation corresponding to the component of the electro-mechanical system.
10. 10 . The method of claim 1 , wherein the electro-mechanical system includes multiple components including a first component and a second component, wherein the processor is configured to determine a first causal FIR filter for estimating velocities of the first component and a second causal FIR filter for estimating velocities of the second component, wherein coefficients of the first causal FIR filter differ from coefficients of the second causal FIR filter.
11. 11 . The method of claim 10 , wherein the electro-mechanical system includes a robot, wherein the multiple components include multiple joints of the robot.
12. 12 . The method of claim 11 , wherein one of the multiple components of the electro-mechanical system includes a motor moving a load, and wherein the task of the one of the multiple components includes one or a combination of a type of the load and a type of motion trajectory for moving the load.
13. 13 . An electro-mechanical system, comprising: multiple joints including at least a first joint having a first mass and a second joint having a second mass; multiple actuators configured to actuate the multiple joints including a first actuator configured to actuate the first joint and a second actuator configured to actuate the second joint; multiple position sensors for measuring positions of the multiple joints including a first position sensor configured to measure positions of the first joint and a second position sensor configured to measure positions of the second joint; multiple causal finite impulse response (FIR) filters configured to estimate velocities of the multiple joints including a first causal FIR filter configured to estimate velocities of the first joint by updating coefficients of the first causal FIR filter to reduce a difference between a sequence of ground truth velocities of the first joint and estimated velocities produced by filtering the measured positions of the first joint and a second causal FIR filter configured to estimate velocities of the second joint by updating coefficients of the second causal FIR filter to reduce a difference between a sequence of ground truth velocities of the second joint and estimated velocities produced by filtering measured positions of the second joint, wherein the coefficients of the first causal FIR filter differ from the coefficients of the second causal FIR filter, wherein first causal FIR filter and the second causal FIR filter correspond to a task performed by the first joint and the second joint; and a feedback controller configured to submit control commands to the multiple actuators, wherein the control commands are determined based on the measured positions and the estimated velocities of the multiple joints.
14. 14 . The electro-mechanical system of claim 13 , further comprising a machine learning module configured to: excite the electro-mechanical system to collect a sequence of measured positions of the first joint and a sequence of measured positions of the second joint; submit the sequence of measured positions of the first joint and the sequence of measured positions of the second joint to an acausal filter configured to estimate velocities with apriori denoising to receive a corresponding sequence of ground truth velocities of the first joint and the sequence of ground truth velocities of the second joint; determine the coefficients of the first causal FIR filter such that outputs of the first causal FIR filter approximate the sequence of ground truth velocities of the first joint; and determine the coefficients of the second causal FIR filter such that outputs of the second causal FIR filter approximate the sequence of ground truth velocities of the second joint.
15. 15 . The electro-mechanical system of claim 13 , wherein the task corresponds to moving a load, wherein the electro-mechanical system further comprises: a machine learning module configured to determine one or a combination of the coefficients of the first causal FIR filter and the coefficients of the second causal FIR filter in response to a change of the task and/or the load.
16. 16 . The electro-mechanical system of claim 13 , wherein a first size of the first causal FIR filter differs from a second size of the second causal FIR filter.
17. 17 . The electro-mechanical system of claim 13 , wherein at least one of the multiple position sensors is an encoder.
18. 18 . The electro-mechanical system of claim 13 , wherein the task corresponds to control of or estimation corresponding to at least one of the multiple joints.

Description

FIELD OF THE DISCLOSURE The present disclosure relates generally to velocity estimation, and more specifically to systems and methods for task-specific adaptive velocity estimation in automated control and monitoring systems. BACKGROUND Velocity estimation plays an important role in the control and monitoring of many mechanical systems whose state is described by both position variables (angles or distances) as well as velocities (angular or linear). More often than not, dedicated velocity sensors (such as tachometers) are not available, and velocities need to be estimated from position sensors, most commonly rotary encoders. Encoders introduce quantization noise, i.e., errors or discrepancies that arise due to the discrete nature of the position information provided, which depends on the resolution of the encoder. This noise, when incorporated into the determined velocity signal, may end up creating a disturbance in the control signal that can significantly worsen the performance of the controller and even render it unstable. In the context of mechanical systems, stability and accuracy in velocity estimation are essential for effective control and monitoring. Addressing the issues related to quantization noise from encoders is therefore desirable for improving the overall performance of the system. SUMMARY The present disclosure provides systems and methods for task-specific adaptive velocity estimation in automated control and monitoring systems of electro-mechanical systems. Generally, a causal finite impulse response (FIR) filter is used to estimate velocities of one or more components of an electro-mechanical system (such as a robot) based on position measurement data. The causal FIR filter is task-specific, as optimal coefficients of the causal FIR filter may vary by task. The task may be defined by a desired-set point or a motion trajectory of the components and/or by properties of a load moved by the components. While the task is being executed, a position sensor (such as an encoder) collects a sequence of position measurement data defining the movement of the component. The sequence of position measurement data is then processed by the causal FIR filter to determine a sequence of estimated velocities. In some examples, the electro-mechanical systems may include multiple causal FIR filters with different coefficients corresponding to different components of the system. In order to produce accurate estimated velocities, the coefficients of the causal FIR filter are adapted by comparing the estimated velocities to a sequence of ground truth velocities corresponding to both the performed task and the electro-mechanical system being monitored. The coefficients are updated to reduce a difference between the estimated velocities and the ground truth velocities, thereby creating a more accurate causal FIR filter to estimate velocities. In some examples, the ground truth velocities may be captured using a tachometer. These captured ground truth velocities may then be used to train the coefficients of the causal FIR filter. In other examples, the ground truth velocities may be determined by processing the position measurement data with an acausal filter. The acausal filter may be trained offline with measured position and velocity data. Some embodiments of the present disclosure are based on recognizing the advantages and disadvantages of using causal filters for estimating the velocity of the operation of the electro-mechanical components from sequential measurements of the position of these electro-mechanical components. As an advantage, the causal filters are fast, computationally efficient, and can be implemented on inexpensive embedded systems with limited computational capabilities. As a disadvantage, the causal filters that consider only previously measured position data inevitably introduce noise and time lags in their estimations. It is an object of some embodiments to attenuate the disadvantages of the causal filters configured to estimate velocities from a sequence of position data while preserving their advantages. Some embodiments are based on recognizing that causal filters, such as causal finite impulse response (FIR) filters, are designed based on various principles agnostic to the specifics of the task of the operation performed by the electromechanical components. Examples of such task-agnostic filter design include first-order backward difference estimator (BDE), Taylor series expansion (TSE), and the like. The chosen design method affects the characteristics of the filter. Considering the variety of possible uses of causal filters and the specifics of their designs and their operations, deployment of such a task-agnostic filter design should not come as a surprise. However, some embodiments of the present disclosure are based on recognizing, informed by mathematical analysis and simulations, that differently designed causal filters can work better for different tasks. In other words, the