SYSTEMS AND METHOD FOR VOLITIONAL CONTROL OF JOINTED MECHANICAL DEVICE BASED ON SURFACE ELECTROMYOGRAPHY

Abstract

Systems and methods for controlling a weight bearing member having at least one powered joint are provided. A system includes a velocity reference module for receiving myoelectric control signals from a user during a non-weight bearing mode for the powered joint and generating a velocity reference for the powered joint based on the myoelectric control signals. The system further includes a volitional impedance module for generating a torque control signal for actuating the powered joint based at least on the velocity reference.

Description

FIELD OF THE INVENTION

The present invention relates to control of jointed mechanical devices, and more specifically to systems and methods for controlling jointed mechanical devices based on surface electromyography.

BACKGROUND

Although prosthetic knee joints for transfemoral prostheses have traditionally been energetically passive devices, powered, semi-autonomous knee joints have recently started to emerge in the research community and on the commercial market. Typically, passive knee prostheses can only react to mechanical energy imparted by the amputee, while powered knee prostheses have the ability to act independently of mechanical energy from the user. As such, the nature of the user communication with the powered prosthesis and control of the powered prosthesis is substantially different from the control of a traditional, energetically passive prosthesis.

Various methods have been proposed for the control of powered knee prostheses. These approaches typically utilize instrumentation on at least one of the prosthesis or a sound leg. Such instrumentation can include inertial measurement sensors (accelerometers and/or gyroscopes) at the foot, shank or thigh of the prosthesis and/or sound side. Additionally, joint angular position, velocity and torque sensors for ankle, knee and hip joints of the prosthesis and/or sound side can also be used as instrumentation for prosthesis control. Further, ground force detecting load cells or load switches can also be used to detect events such as heel strike or toe-off. This instrumentation is used to form knee joint angle trajectories or impedances for the powered knee prosthesis during activities involving the prosthesis. For example, while standing, walking, or transitioning between sitting and standing.

In general, activities such as standing, walking, or transitioning between sitting and standing all involve physical input and/or energy exchange between the residual limb and prosthesis. Therefore, most conventional methods rely on some form of physical input from the user for communication with the powered knee prosthesis. That is, although the user need not provide the energy for movement, as is the case with traditional dissipative knee prostheses, the user must still provide some physical input that can be measured by instrumentation on the prosthesis and/or sound leg. Such physical inputs include measuring weight bearing on the prosthesis, torque and/or acceleration from the affected-side hip joint, movement of the sound-side leg, to name a few.

An important class of movement, however, which does not involve any significant physical input from the user, is the task of non-weight-bearing or volitional control of knee movement while sitting or standing. That is, people regularly shift their body while sitting, which involves significant movement of the knee joints. Such movement has both physiological and practical purposes. Regarding the former, weight shifting during sitting is known to play an important role in ensuring healthy circulation of blood in weight-bearing tissues during sitting. Regarding the latter, sitting in confined areas, such as in automobiles, airplanes, theatres, and classrooms, often requires shifting of body position (particularly of the knee joints) in order to accommodate a particular ergonomic space and/or the movement of other individuals into or out of that space. Such movement is referred to herein as volitional control of the knee joint during non-weight-bearing activity. Note that such volitional control is also useful in non-weight-bearing standing, such as when flexing the knee to look at the bottom of a shoe, or when placing the foot on an elevated surface (such as a chair) to tie or untie, or don or doff a shoe. In the case of a traditional, energetically passive prosthesis, an amputee typically achieves “volitional” control functionality by manipulating the prosthetic knee leg with his or her hands.

SUMMARY

The various embodiments of the invention concern systems and methods for controlling jointed mechanical devices based on surface electromyography. In a first embodiment, a myoelectric controller for a weight bearing member having at least one powered joint is provided. The myoelectric controller includes a velocity reference module for receiving myoelectric control signals from a user during a non-weight bearing mode for the powered joint and generating a velocity reference for the powered joint based on the myoelectric control signals. The controller further includes a volitional impedance module for generating a torque control signal for actuating the powered joint based at least on the velocity reference.

In a second embodiment of the invention, a method for controlling at least one powered joint in a weigh bearing member is provided. The method includes the step of receiving myoelectric control signals from user during a non-weight bearing mode for the powered joint. The method also includes generating a velocity reference for the powered joint based on the myoelectric control signals. The method further includes generating a torque control signal for actuating the powered joint based at least on the velocity reference.

In a third embodiment of the invention, a jointed mechanical device is provided. The device includes a weight bearing member comprising at least one powered joint. The device also includes a controller for actuating the powered joint. In the device, the controller is configured for actuating the joint in at least one of a semi-autonomous weight bearing mode and a non-weight bearing mode for actuating the powered joint responsive to myoelectric control signals.

In a fourth embodiment of the invention, a computer-readable medium is provided, storing instructions for controlling a computing device to control a powered joint. The instructions include instructions for receiving myoelectric control signals from user during a non-weight bearing mode for the powered joint, generating a velocity reference for the powered joint based on the myoelectric control signals, and generating a torque control signal for actuating the powered joint based at least on the velocity reference.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 shows a block diagram of a myoelectric volitional impedance controller for controlling a powered knee joint in accordance with an embodiment.

FIG. 2 shows a block diagram of a myoelectric volitional impedance controller for controlling a powered knee joint and a powered ankle joint in accordance with an embodiment.

FIG. 3A is an x-y plot of extension and flexion reference signals for a first amputee subject showing classification using QDA and LDA methods in accordance with the various embodiments.

FIG. 3B is an x-y plot of extension and flexion reference signals for a second amputee subject showing classification using QDA and LDA methods in accordance with the various embodiments.

FIG. 3C is an x-y plot of extension and flexion reference signals for a third amputee subject showing classification using QDA and LDA methods in accordance with the various embodiments.

FIG. 4A is an x-y plot of actual measurements and PCA projections of extension reference signals for the first amputee subject in FIG. 3A.

FIG. 4B is an x-y plot of actual measurements and PCA projections of flexion reference signals for the first amputee subject in FIG. 3A.

FIG. 5 is a schematic illustration of an exemplary powered transfemoral prosthesis that can be configured for using a control system in accordance with the various embodiments.

FIG. 6A-6D shows x-y plots of EMG-controlled powered prosthesis knee position for trajectories A-D of the third amputee subject in FIG. 3C.

FIG. 7A-7D shows x-y plots of sound-side knee position for trajectories A-D of the third amputee subject in FIG. 3C.

DETAILED DESCRIPTION

The present invention is described with reference to the attached figures, wherein like reference numerals are used throughout the figures to designate similar or equivalent elements. The figures are not drawn to scale and they are provided merely to illustrate the instant invention. Several aspects of the invention are described below with reference to example applications for illustration. It should be understood that numerous specific details, relationships, and methods are set forth to provide a full understanding of the invention. One having ordinary skill in the relevant art, however, will readily recognize that the invention can be practiced without one or more of the specific details or with other methods. In other instances, well-known structures or operations are not shown in detail to avoid obscuring the invention. The present invention is not limited by the illustrated ordering of acts or events, as some acts may occur in different orders and/or concurrently with other acts or events. Furthermore, not all illustrated acts or events are required to implement a methodology in accordance with the present invention.

As described above, in a traditional, energetically passive prosthesis, an amputee can achieve volitional control functionality by manipulating the prosthetic knee leg with his or her hands. However, since a powered knee prosthesis has the capability to move itself, such artificial manipulation should not be required for volitional movement of the knee joint. Nonetheless, since such volitional movements do not involve significant physical input from the amputee, conventional control approaches do not provide an effective means of communication with the prosthesis for this purpose.

In view of the limitations of conventional control approaches and conventional prostheses, the various embodiments provide a new method for the volitional control of the knee joint during non-weight-bearing activities. More generally, the various embodiments provide systems and methods controlling jointed mechanical devices, such as prostheses and orthoses, during non-weight bearing activities based on a volitional impedance control framework.

For example, in the case of a leg prosthesis, this allows a transfemoral amputee to control the motion of a powered knee prosthesis during non-weight-bearing activity (e.g., while sitting.). The control is based on an impedance framework wherein the joint exhibits programmable joint stiffness and damping characteristics. Knee movement is provided by commanding the joint stiffness equilibrium angle. The time rate of change of this angle (which is the desired angular velocity set-point) is provided by measurement of the surface EMG, using a pair of surface electromyogram (EMG) electrodes. In one embodiment, the electrodes can be on the ventral and dorsal aspects of the thigh. For example these electrodes can be integrated into the amputee's socket interface so as to measure the surface EMG of the hamstring and quadriceps muscle groups. However, rather than directly associate the hamstring EMG with knee flexion and the quadriceps with knee extension, which would require the user to artificially isolate contraction of these muscle groups, the various embodiments incorporate a combination of pattern classification and principal component projection to align the measured EMG with the user's desire to flex or extend the knee joint. The resulting control approach provides trajectory tracking performance close to that of intact knee joints, thus providing an approach for effective control of knee joint motion during non-weight-bearing activity. Further, this approach can be integrated with existing impedance-based weight-bearing controllers for standing, walking, and transitioning between sitting and standing. For example, this approach can be integrated into the controller described in U.S. patent application Ser. No. 12/427,384 to Goldfarb et al, filed Apr. 21, 2009, the contents of which are herein incorporated in their entirety.

Although exemplary embodiments will be described primarily with respect to providing volitional control for a prosthesis including at least a powered knee, the various embodiments are not limited in this regard. Rather, the framework described herein can be used for volitional control of any type of powered joint in a prosthesis or an orthosis.

The use of surface EMG for the control of lower limb prostheses and orthoses has been widely investigated. In the case of passive knee prostheses, one existing approach provides a prosthesis with an electrically activated knee flexion lock that uses surface EMG from the residual limb of a transfemoral amputee to trigger the engagement and disengagement of the lock. A similar approach includes a computer-controllable passive knee prosthesis based on an electrically modulated brake, and utilizes surface EMG from three sites on the residual limb of a transfemoral amputee for gait mode recognition, which in turn was used to switch the prosthesis into the appropriate gait mode. More recently, surface EMG from multiple electrodes on transfemoral amputees has been utilized to classify movement intents while walking. However, with regard to using EMG for the real-time control of a powered knee prosthesis, only limited investigation into the use of surface EMG from the quadriceps and hamstrings to control the motion of a hydraulically actuated powered knee prosthesis during walking has occurred. Further, such research concluded that use of such an approach during gait would be challenging, due in part to difficulty in obtaining reliable EMG measurement, “due to noise pick up and movement artifact.”

Other conventional control methodologies using surface EMG measured from the lower leg have been generally directed to the control of powered ankle joints in transtibial prostheses or control of powered joints in ankle-foot-orthoses (AFOs). With regard to the former, one approach provides for using a real-time state-based controller for the powered ankle based on physical input (rather EMG input) from the user and which utilizes EMG measured from the lower leg to switch between gait modes. With regard to powered AFOs, one approach uses EMG measured from the lower leg to control the assistive pressure in a pneumatically actuated AFO. Additionally, a control system for an assistive exoskeleton with powered hip and knee joints has been proposed, in which the assistive torque from the exoskeleton is proportional to the measured EMG from the associated flexion or extension muscle group. However, none of these approaches utilize EMG for the volitional control of knee joint motion in a powered knee prosthesis.

Volitional Control of Powered Knee

A. Volitional Control Structure

In a first exemplary embodiment, a control framework is provided for volitional control of the knee with a joint output impedance similar to that of the native limb. As such, rather than using the measured EMG to prescribe joint torque, angle, or angular velocity, the presented framework utilizes measured EMG to prescribe the angular velocity of an equilibrium point of joint impedance that consists of the combination of a joint stiffness and damping. In this manner, the knee moves to a desired position with a joint output stiffness and damping prescribed by the controller, thus presumably moving in a more natural manner (relative to a high-output-impedance position controller), and resulting in a more natural interaction between the user, prosthesis, and environment.

The structure of the proposed volitional controller is shown in FIG. 1. FIG. 1 shows a block diagram of a myoelectric volitional controller 100 for controlling a powered knee joint in a prosthesis 102 based on EMG signals from a user 104 in accordance with an embodiment. In this controller, a real-time intent recognizer, such as the one described in U.S. patent application Ser. No. 12/427,384 to Goldfarb et al, filed Apr. 21, 2009, or in Varol, H. A., Sup, F., and Goldfarb, M. Multiclass Real-Time Intent Recognition of a Powered Lower Limb Prosthesis. IEEE Transactions on Biomedical Engineering, vol. 57, no. 3, pp. 742-751, 2009 would be used to switch between this (volitional) controller and other weight-bearing control structures.

With reference to FIG. 1, the controller 100 operates as follows. First, during a non-weight bearing activity, based on an intent recognizer (not shown), EMG signals (EMG₁, EMG₂) are received by controller 100 from the residual limb of the user 104. Thereafter, a pre-processing module 106 processes the EMG signals. The pre-processed signals are then used for generation of reference velocity (i.e., the joint angular velocity reference, ω_emg). As shown in FIG. 1, ω_emg is generated by a velocity reference generation module 108 based on the pre-processed the EMG signals, the current angle for the joint (θ_k), the derivative or rate of change of angle for the joint (r), and the user's intent to flex or extend the knee. The intent can be obtained from a flexing extending classification module 110. Thereafter, an equilibrium point joint angle θ_emg can be obtained from ω_emg using conversion module 112. The equilibrium point joint angle, θ_k, and {dot over (θ)}_k can then be used in a volitional impedance controller module 114 to generate a joint torque command (t) for the prosthesis 102 to cause motion of the knee joint. Additionally, θ_k and {dot over (θ)}_k are updated based on the torque command. The control process then repeats. It should be noted though that since θ_k and {dot over (θ)}_k are the measurements of the knee angle and velocity of the prosthesis, the update is not done computationally, rather it is a physical process. The operation of these various modules is described below in greater detail.

B. Volitional Impedance Controller

In the various embodiments, EMG is used to generate an angular velocity command (as is commonly the case in upper extremity myoelectric control) rather than a position command, so that the user contracts the residual limb musculature only to move the joint and can relax when maintaining any given knee joint angle. Specifically, the joint torque command at module 114 can be given by a model mimicking the behavior of a spring and dashpot element. For example, one model in accordance with the various elements can be:

τ=k(θ−θ_emg)+b{dot over (θ)}  (1)

where the equilibrium point θ_emg is given by module 112 using

$\begin{array}{cc}{\theta }_{\mathrm{emg}}={\theta }_o+{\int }_t{\omega }_{\mathrm{emg}}\mathrm{dt}& \left(2\right)\end{array}$

where k is the prescribed joint stiffness, b is the prescribed joint damping coefficient, θ is the knee joint angle, and θ_o is the initial angle when the control system switches to the volitional (non-weight bearing) controller and ω_emg is the angular velocity reference generated from the quadriceps and hamstring EMG, as described in the following section.

C. Reference Velocity Generation

The impedance controller 114 utilizes the measured surface EMG from the quadriceps and hamstring groups to generate a joint angular velocity reference, ω_emg, to drive the joint angular impedance equilibrium point, θ_emg, and thus to drive the motion of the knee. One method for doing so at module 108 would be to use

$\begin{array}{cc}{\omega }_{\mathrm{emg}}=\{\begin{array}{cc}{k}_he_h& \mathrm{if}e_h\ge e_q\\ -k_qe_q& \mathrm{otherwise}\end{array}& \left(3\right)\end{array}$

where e_h and e_q represent the measured (i.e., rectified and filtered) EMG from the hamstring and quadriceps muscles, respectively, and k_h and k_q are simple gains. Equation (3) also assumes that an appropriate dead-band is applied to the measured EMG, to avoid “jitter” in the angular velocity reference command. Equation (3) is similar to the method used for the control of myoelectric upper extremity prostheses.

However, as shown below, use of equation (3) provided only marginal performance in the proposed volitional controller. Specifically, as described subsequently (and indicated in FIGS. 3A-3C), two of the three amputee subjects on which the approach was implemented demonstrated a significant degree of co-contraction when attempting to contract either the hamstrings or quadriceps in an isolated manner. With sufficient training, these subjects could possibly be trained to avoid co-contraction. Co-contraction, however, is a natural neuromuscular response (particularly in the lower limb musculature). As such, in an effort to render the proposed controller as natural as possible, in the various embodiments, the controller is trained to properly interpret co-contraction, rather than train the subjects to avoid it. Therefore, as indicated in the control structure of FIG. 1 and described below, the controller 100 first utilizes pattern classification to classify the user's intent with regard to flexion or extension of the knee, then utilizes a projection operator to extract the desired magnitude of the joint angular velocity reference from the measured EMG data.

D. Flexion-Extension Classification

As described above, rather than train the subjects to avoid co-contraction while commanding flexion or extension of the knee, the various embodiments utilize a pattern classification approach to distinguish user intent to flex or extend the knee. In one embodiment, module 110 can be implemented using a quadratic discriminant analysis (QDA) classifier to distinguish between the user's intent to flex or extend. A linear discriminant analysis (LDA) classifier can also applied to the classification problem in other embodiments, although the QDA was chosen due to improved classification accuracy (based on the mean accuracy obtained with a five-fold cross-validation for each subject), and because the QDA is not significantly more complex (or computationally expensive) than the LDA classifier. Specifically, QDA uses the quadratic decision boundary of the form c₁+c₂e_h+c₃e_q+c₄e²_h+c₅e_he_q+c₆e²_q=0 to classify the sample consisting of the processed EMG data from the two channels, e_h and e_q, to the extension (ω_E) and flexion (ω_F) classes where the coefficients c_i, i=1, 2, . . . , 6, are generated during the training of the QDA classifier. Details of the LDA and QDA methods can be found in several pattern classification references. Further, this embodiment utilizes a database of EMG (versus intent) data to parameterize the flexion/extension classifier, as described below.

E. EMG Measurement and Preprocessing

In the various embodiments, the electrodes can be implemented in various ways. For example, as described above, surface EMG electrodes can be embedded into the prosthesis socket. In another implementation, separate surface electrodes can be placed on the amputee to acquire EMG signals from the residual quadriceps and hamstring muscles of the amputee subjects. To improve acquisition of such signals, the signals from each muscle group can amplified, filtered, and/or rectified at module 106. In other words, the EMG preprocessing attempts to discern an envelope of the raw EMG signal.

In the various embodiments, each EMG signal is acquired from a single bipolar electrode. Such signals are generally very small in magnitude and can have both positive and negative values when the muscle contracts. An instrumentation amplifier can then be used to increase the voltage levels of these signals. Thereafter, high pass filtering can be applied to remove the baseline noise and rectification can be done to remove the negative values. Finally, the signal can be low pass filtered to create an envelope of the signal. This way, the muscle contraction EMG signals will be converted to a unidirectional multilevel signal in expense of some phase delay due to the filtering. In other words, the preprocessing is done to convert the noisy raw EMG data to a less noisy form that is more suitable for controls and pattern recognition.

For example, in one embodiment, the signal can be processed using an instrumentation amplifier with a gain of 200 and filtered using an analog second order low pass filter with 5 Hz cutoff frequency. The filtered signals can then be digitized for use. For example, using a computer running MATLAB Real Time Workshop with a digital-to-analog converter card and operating at 1000 Hz sampling frequency. The digital signals can then be further processed using additional signal processing. For example, signals can be processed using a first order high-pass filter with 20 Hz cutoff frequency, a rectifier, and a first order low-pass filter with 2 Hz cutoff frequency.

F. EMG Intent Database Generation

Classifier training database generation can be performed by recording EMG data associated with an amputee. For example, in one embodiment, a training database can be generated for an amputee by recording 100 seconds of EMG data for knee flexion and 100 seconds for knee extension. For ease of the subject, a one-minute rest in between the recordings can be provided. Thus the entire training session can be configured to last less than five minutes. To generate a complete set of training data for each flexion/extension class, each subject can be asked to visualize extending the knee on the amputated side at 0, 25, 50, 75 and 100 percent of full effort, several times for durations ranging from 1 to 5 seconds, over the total data collection period of 100 seconds at a 100 Hz sampling frequency. The extension data can be recorded first, followed by a rest period of approximately one minute, followed by the same procedure for flexion data. All EMG data can be normalized into the interval [0, 1]. The data can be additionally thresholded at 20% maximum effort, such that samples in the interval [0, 0.2] are effectively removed from the database and in order to mitigate baseline EMG noise and muscular tonicity. Based on this thresholded database, the QDA classifier can be parameterized to classify each subject's preprocessed EMG as intent to either flex or extend the knee joint.

G. Reference Velocity Magnitude

The QDA essentially provides a probabilistic optimal separation boundary of the EMG data to the flexion and extension classes. Within a given class (in this case flexion or extension), the “magnitude” of the data is the projection along the principal axis of that class. In the control approach described herein, this projection can be generated via principal component analysis (PCA), which essentially projects the two-dimensional EMG data along a principal (either flexion or extension) axis. Using the data belonging to each class, two 2×2 PCA projection matrices W_E and W_F can be computed. In the real-time implementation, one of these projection matrices can be used to extract the “magnitude” information, based on the result of QDA classification as follows:

$\begin{array}{cc}{\left[x_px_s\right]}^T=\{\begin{array}{cc}{W_E\left[x_hx_q\right]}^T& \mathrm{if}{\omega }_E\\ {W_F\left[x_hx_q\right]}^T& \mathrm{if}{\omega }_F\end{array}& \left(4\right)\end{array}$

The magnitude of the angular velocity reference for the joint impedance set-point, ω_emg, can therefore be the PCA-based projection of the two-dimensional EMG data along the principal axis of either the flexion or extension data. Details of PCA can be found in several references. The projected EMG data can be scaled between zero and maximum reference velocity to generate the desired angular velocity reference. The maximum reference velocity can be determined as the maximum reasonable angular velocity command for volitional control of the knee joint.

In contrast with (3), which obtains a reference angular velocity (for the volitional control impedance set-point) by projecting data along a hamstring/quadriceps set of measurement axes, the approach combining QDA classification with PCA projection of the two-dimensional EMG data establishes a probabilistically optimal linear transformation from a hamstring/quadriceps set of axes to a flexion/extension set of axes (based on the training dataset). As such, the subject need not be trained to isolate the contraction of individual muscle groups, but rather is free to co-contract the hamstring and quadriceps groups in a natural manner when intending knee flexion or extension.

II. Volitional Control of Powered Knee and Powered Ankle

A. Volitional Control Structure

Although the volitional control structure above describes how to provide volitional control of a powered knee joint, volitional control of a powered knee and a powered joint is desirable in many circumstances. In particular, volitional control of both the knee and ankle allows more natural motion and provides the user the option to manipulate the foot. Accordingly, in some embodiments, a control framework can be provided that is intended to provide volitional control of both the knee and ankle joints with a joint output impedance similar to that of the native limb. As such, rather than using the measured EMG solely to prescribe joint torque, angle, or angular velocity, the presented framework can utilize measured EMG to prescribe the angular velocity of an equilibrium point of joint impedance that consists of the combination of a joint stiffness and damping. In this manner, the knee can move to a desired position with a joint output stiffness and damping prescribed by the controller, thus presumably moving in a more natural manner (relative to a high-output-impedance position controller), and resulting in a more natural interaction between the user, prosthesis, and environment.

The structure of the volitional controller for knee and ankle joints is shown in FIG. 2. FIG. 2 shows a block diagram of a myoelectric volitional controller 200 for controlling a powered knee joint and a powered ankle joint in accordance with an embodiment. In many respects, the controller 200 operates in a substantially similar fashion to controller 100 in FIG. 1. That is, during a non-weight bearing activity, based on an intent recognizer (not shown), EMG signals (EMG₁, EMG₂) are received by controller 200 from the residual limb of the user 204. Thereafter, a pre-processing module 206 processes the EMG signals. The pre-processed signals are then used for generation of reference velocities for the knee or the ankle joints (i.e., one of joint angular velocity references, ω_{knee_emg} and Cθ_{ankle_emg}, respectively). As shown in FIG. 2, the one of θ_{knee_emg} and ω_{ankle_emg} is generated by a velocity reference generation module 208 based on the pre-processed EMG signals, the user's intent to flex or extend the knee, a user selection of a joint, and a corresponding current angle for the joint to be controlled (θ_kne or θ_ank), the derivative or rate of change of angle for the joint to be controlled ({dot over (θ)}_kne, or {dot over (θ)}_ank). The intent can be obtained from a flexing extending classification module 210. The user selection can be based on detection of co-contraction or a “twitch” using module 211. Thereafter, an equilibrium point joint angle for the one of the knee (θ_{kne_emg}) or the ankle (θ_{ank_emg}) can be obtained from a corresponding one of ω_{knee_emg} and ω_{ankle_emg} using conversion module 212. The one equilibrium point joint angle can then be used in a volitional impedance controller module 114 to generate a joint torque command (t) for the prosthesis 102 to cause motion of the knee or ankle joint. Additionally, the resulting values for θ_kne, θ_ank, {dot over (θ)}_kne, and {dot over (θ)}_ank are updated. The control process then repeats. The operation of these various modules is described below in greater above and below, as necessary.

B. Volitional Impedance Controller

Module 214 here operates in a manner substantially similar to that of module 114 in FIG. 1. Again, it is noted that the EMG is used to generate an angular velocity command (as is commonly the case in upper extremity myoelectric control) rather than a position command, so that the user contracts the residual limb musculature only to move the joint and can relax when maintaining any given knee joint angle. Specifically, the joint torque command at module 214 is given by:

τ_kne=k_kne(θ_kne−θ_{emg_kne})+b_kne{dot over (θ)}_kne

τ_ank=k_ank(θ_ank−θ_{emg_ank})+b_ank{dot over (θ)}_ank  (5)

where the knee and ankle equilibrium points θ_{kne_emg} and θ_{ank_emg} are given by

$\begin{array}{cc}{\theta }_{\mathrm{kne\_emg}}={\theta }_{\mathrm{o\_kne}}+{\int }_t{\omega }_{\mathrm{kne\_emg}}\mathrm{dt}{\theta }_{\mathrm{ank\_emg}}={\theta }_{\mathrm{o\_ank}}+{\int }_t{\omega }_{\mathrm{ank\_emg}}\mathrm{dt}& \left(6\right)\end{array}$

where k_kne and k_ank are the prescribed joint stiffnesses, b_kne and b_ank are the prescribed joint damping coefficient, θ_kne and θ_ank are the knee and ankle joint angles, and θ_{o_kne} and θ_{o_ank} are the initial knee and ankle angles when the control system switches to the volitional (non-weight bearing) controller. ω_{kne_emg} and ω_{ank_emg} are the knee and ankle angular velocity references generated from the quadriceps and hamstring EMG, as described in the following section.

C. Active Joint Selection Using Twitch

In this embodiment, two unidirectional EMG channels are used to generate two bidirectional joint velocity references, one for the ankle joint and one for the knee joint of the prosthesis. In one embodiment, one bidirectional signal for one of the joints is generated at a single instant. In order to achieve this, a short duration co-contraction (twitch) of the both EMG electrode sites can be utilized to select the active joint. A twitch can be detected when the filtered and rectified EMG signals on both channels exceed a threshold for a short duration of time. Once a twitch is detected the active joint of the prosthesis can be toggled. One of the two tactors (e.g. cellphone vibration motor) on the residual limb (or in the socket as shown in FIG. 5 below) can then be activated for a short duration to notify the user which joint is active.

D. Flexion-Extension Classification

Once the active joint is selected, the user intent to flex or extend can be detected using a pattern classification algorithm in module 210, as described above with respect to module 110 in FIG. 1. For example, such algorithms can include linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), support vector machines (SVM) or artificial neural networks (ANN). In order to generate the flexion extension classification boundaries (or functions) a database of different intensity flexion and extension EMG data from hamstring and quadriceps muscles need to be collected, as described above. This data can then be used to train the pattern classifiers.

D. Reference Velocity Magnitude

Once the user intent to flex and extend is inferred, the joint velocity reference magnitude can be obtained at module 208 as a function of the filtered EMG signals. One possible way to generate the velocity references might be to use principal component analysis to project the two dimensional filtered EMG signals to the one dimensional principal component.

III. Experimental Implementation

A. EMG-Based Reference Velocity Generation

The proposed volitional knee joint controller of FIG. 1 was implemented on three transfemoral amputee subjects. The subjects were all male, between the ages of 20 and 60, and between 3 months and 4 years post amputation. Two of the subjects were unilateral transfemoral amputees, while one subject (subject 3) was a bilateral amputee, with a transfemoral amputation on one leg and a transtibial on the other. In all cases, all subjects were characterized by a prosthetic knee on one limb and an intact knee on the other.

FIGS. 3A-3C shows the EMG intent databases corresponding to each subject. In particular, FIG. 3A is an x-y plot of extension and flexion reference signals for a first amputee subject (subject 1) showing classification using QDA and LDA methods in accordance with the various embodiments. FIG. 3B is an x-y plot of extension and flexion reference signals for a second amputee subject (subject 2) showing classification using QDA and LDA methods in accordance with the various embodiments. FIG. 3C is an x-y plot of extension and flexion reference signals for a third amputee subject (subject 3) showing classification using QDA and LDA methods in accordance with the various embodiments.

These databases, as described above, correspond to 100 seconds of flexion data at various degrees of (muscular) effort, and 100 seconds of extension data, also at various degrees of effort. Note that the x_q axis represents the measured, preprocessed, normalized, and thresholded EMG for the quadriceps group, while the x_h axis represents the EMG measured for the hamstring group. As seen in FIGS. 3A-3C, two of the three subjects (subjects 1 and 3) demonstrated a significant amount of muscular co-contraction when intending volitional movement of the prosthetic knee. Interestingly, subject 1 primarily demonstrated significant co-contraction during intent to extend the knee, while subject 3 primarily demonstrated significant co-contraction during intent to flex the knee. For all subjects, the LDA and QDA boundaries between classes along with the pseudo-classification boundary described by (3) are shown in the figures. Recall that, based on a five-fold cross-validation of classification accuracy, QDA classification in general provided higher classification accuracies, and therefore was used in the control experiments to classify intent to flex or extend the knee. Specifically, the mean accuracies of the classifiers over 5 CV-fold for each of the three subjects are 0.99, 0.80 and 0.86 for the LDA and 1.0, 0.86 and 0.90 for the QDA. Note that, particularly in the cases of subjects 1 and 3, the simple thresholding approach (described by (3)) entails a considerable amount of erroneous “classification” of intent, even in the case of large amplitude EMG (x_i>0.3). In contrast, the QDA classification boundaries entail little to no classification error, particularly in large amplitude EMG.

Once intent to flex or extend the knee is known, the magnitude of the angular velocity for the impedance set-point is obtained by projecting the corresponding data point onto its principal axis via PCA. A representative example of the corresponding PCA projections for subject 1 is shown in FIG. 4. In the figure, the x_p axis corresponds to the PCA projection of the flexion and extension data along the principal component of that data. As such, the angular velocity for the impedance set-point of the volitional knee joint controller can be given by:

$\begin{array}{cc}{\omega }_{\mathrm{emg}}=\{\begin{array}{cc}0& \mathrm{if}{\omega }_{F}\mathrm{and}\left(x_P<\gamma \right)\\ \alpha \left(\frac{x_p-\gamma }{1-\gamma }\right)& \mathrm{if}{\omega }_F\mathrm{and}\left(x_P\ge \gamma \right)\\ 0& \mathrm{if}{\omega }_E\mathrm{and}\left(x_P<\gamma \right)\\ -\alpha \left(\frac{x_p-\gamma }{1-\gamma }\right)& \mathrm{if}{\omega }_E\mathrm{and}\left(x_P\ge \gamma \right)\end{array}& \left(7\right)\end{array}$

where α is the maximum desired set-point velocity (corresponding to maximum muscular effort), γ is the value at which the normalized EMG is thresholded (in this case γ=0.2), x_p is the PCA projection along the principal axis. This is shown in FIGS. 4A and 4B.

FIG. 4A is an x-y plot of actual measurements and PCA projections of extension reference signals for the first amputee subject in FIG. 3A. FIG. 4B is an x-y plot of actual measurements and PCA projections of flexion reference signals for the first amputee subject in FIG. 3A. For the actual samples, x_q and x_h denote the normalized EMG signals for the quadriceps and hamstrings muscles, respectively. For the PCA projections, x_p and x_s denote the first principal and second principal components, respectively.

B. Volitional Trajectory Tracking of a Powered Knee Prosthesis

The volitional knee controller of FIG. 1 was implemented on each of the three amputee subjects with the powered transfemoral prosthesis shown in FIG. 6 and described in detail in U.S. patent application Ser. No. 12/427,384 to Goldfarb et al, fil_ed Apr. 21, 2009, the contents of which are herein incorporated in their entirety. The prosthesis used in these experiments _also co_ntains a powered ankle, although the ankle was not explicitly commanded in these experiments, but rather remained in a “neutral” configuration. In order to characterize the effectiveness of the volitional controller for purposes of moving the knee joint, an experiment was developed which required each subject to track various types of knee joint angle movements. During these experiments, each amputee was presented with a computer monitor that showed in real-time a desired knee angle, along with the knee angle of the powered prosthesis, as measured by the joint angle sensor on the prosthesis.

Prosthesis sockets with embedded EMG electrodes for each subject were not available for these experiments. Normally, such EMG electrodes would be disposed similarly to the configuration illustrated in FIG. 5 in order to correspond with locations of the residual portions of the hamstring and/or quadriceps on an amputee. However, since the various embodiments of the volitional controller are intended for non-weight-bearing activity such as sitting, the subjects did not wear the powered prosthesis during the knee control experiments, but rather the subjects were seated in a chair and the powered knee prosthesis was mounted to a bench immediately next to the subject. The prosthesis was mounted in an orientation that was consistent with the seated position of the subjects.

Aside from the QDA and PCA parameters extracted from the EMG intent database, all subjects utilized the same set of volitional control parameters for the powered prosthesis. Specifically, the stiffness of the impedance controller was selected as k=1.0 Nm/deg, the damping as b=0.01 Nm/deg/s, the maximum set-point velocity α=50 deg/s. These parameters were selected experimentally to provide an acceptable bandwidth of motion, while maintaining a natural appearance of motion and a stable interaction with obstacles in the environment (e.g., the leg of a chair).

In order to characterize volitional control of various types of motion, four different desired trajectories were constructed (referred herein as trajectories A through D). The trajectory A joint angle tracking task consisted of set point trajectories requiring the subject to quickly change the knee angle in 8 to 45 degree increments and to hold it for 5 to 10 seconds. Trajectory B consisted of sloped trajectories, which were intended to measure the subject's ability to move the prosthesis at different constant velocities. Trajectories C and D consisted of sinusoidal waves at 0.2 and 0.33 Hz, respectively (i.e., five-second and three-second periods, respectively), which were intended to measure the subject's ability to move the leg up and down smoothly at continuously varying velocities. Trajectories A and B lasted for a total duration of 160 and 180 seconds, respectively, while trajectories C and D lasted for a total duration of 60 seconds each.

For each amputee subject, three sessions of experiments were conducted, each on a different day, with each successive session approximately one week apart. During the experimental sessions, the amputee spent approximately one hour practicing the tracking of the four trajectories (A through D), during which the various trajectories were presented to the amputee in an arbitrary order. After completion of the third session (i.e., after approximately one hour of practice in the third session), the subject's performance was evaluated in a single set of performance tests, consisting of one trial each of trajectories A through D. Representative trajectory tracking performance data corresponding to subject 3, whose average performance was between that of subjects 1 and 2, is shown in FIG. 6. FIGS. 6A-D shows x-y plots of EMG-controlled powered prosthesis knee position for trajectories A-D, respectively of the third amputee subject in FIG. 3C.

The root-mean-square (RMS) trajectory tracking error for all amputee subjects for each of the four trajectories is summarized in Table I. As seen in the table, the average RMS tracking error across all subjects and all trajectories was 6.2 deg.

TABLE I
RMS ERROR FOR EMG CONTROL OF POWERED KNEE
	Subject 1	Subject 2	Subject 3	EMG Control
	EMG	EMG	EMG	Avg.
Trajectory A	6.8	8.2	8.0	7.7
Trajectory B	2.5	3.9	3.7	3.4
Trajectory C	4.4	7.2	5.3	5.6
Trajectory D	8.4	8.3	7.1	7.9
Subject Avg.	5.5	6.9	6.0	6.2

B. Comparison to Intact Knee Trajectory Tracking

In order to provide context for the trajectory tracking data summarized in Table I, corresponding experiments were conducted to assess the ability of each amputee to track the same set of knee joint angle trajectories with his sound knee. These experiments were conducted in a single session, since familiarization with the prosthesis and volitional impedance controller was not necessary (i.e., each subject was already quite familiar with the movement control of his sound knee). As such, each subject spent approximately 15 minutes practicing each set of trajectories, until each was comfortable with his ability to track the trajectories. Once sufficiently comfortable, each subject's performance was evaluated in a single set of performance tests, consisting of one trial each of trajectories A through D. Movement of the subjects' sound knee was measured by using a knee brace instrumented with a goniometer. The knee brace did not impose any significant constraints on knee movement. Representative data corresponding to subject 3 (whose prosthetic side data is shown in FIG. 6) is shown in FIGS. 7A-D. FIGS. 7A-D shows x-y plots of sound-side knee position for trajectories A-D, respectively, of the third amputee subject in FIG. 3C.

The RMS trajectory tracking error for sound side knee angle tracking for all subjects for each of the four trajectories is summarized in Table II.

TABLE II
RMS ERROR FOR VOLITIONAL CONTROL OF INTACT KNEE
	Subject 1	Subject 2	Subject 3
	Sound	Sound	Sound	Sound Side Avg.
Trajectory A	6.1	6.8	7.6	6.8
Trajectory B	1.4	1.8	3.1	2.1
Trajectory C	4.6	6.1	6.4	5.7
Trajectory D	4.5	6.4	7.7	6.2
Subject Avg.	4.2	5.3	6.2	5.2

As seen in Table II, the average RMS tracking error across all subjects and all trajectories for sound knee tracking was 5.2 deg. Recall from Table I that the average RMS tracking error across all subjects and all trajectories for the EMG-based prosthesis knee tracking was 6.2 deg, thus indicating a difference in tracking error between the prosthetic and intact knee joints of one degree. As such, as indicated collectively by the tracking data, the ability of the amputee to control non-weight-bearing knee joint motion of the powered prosthesis (with the EMG-based impedance controller) is nearly as good as their ability to control non-weight-bearing knee joint motion in their intact knee. Further, the performance differences were fairly invariant with respect to movement type. Specifically, the average RMS errors for trajectory A (steps) were 7.7 deg and 6.8 deg, respectively, for the prosthetic and intact joints, and thus the difference in average error was 0.9 deg. The average RMS errors for trajectory B (ramps) were 3.4 deg and 2.1 deg, respectively, for the prosthetic and intact joints, and thus the difference in average error was 1.3 deg. For trajectory C (the slower sinusoid), the average RMS errors were 5.6 deg and 5.7 deg, respectively, for the prosthetic and intact joints, and thus the tracking performance for the slower sinusoid was essentially the same for the prosthetic and intact joint control. Finally, for trajectory D (the faster sinusoid), the average RMS errors were 7.9 deg and 6.2 deg, respectively, for the prosthetic and intact joints, and thus the prosthesis controller demonstrated 1.7 deg more error on average than the intact joint.

The resulting control approach shows that the resulting volitional control provides trajectory tracking performance close to that of their respective intact knee joints, thus indicating that the approach provides effective control of knee joint motion during non-weight-bearing activity.

While various embodiments of the present invention have been described above, it should be understood that they have been presented by way of example only, and not limitation. Numerous changes to the disclosed embodiments can be made in accordance with the disclosure herein without departing from the spirit or scope of the invention. Thus, the breadth and scope of the present invention should not be limited by any of the above described embodiments. Rather, the scope of the invention should be defined in accordance with the following claims and their equivalents.

Although the invention has been illustrated and described with respect to one or more implementations, equivalent alterations and modifications will occur to others skilled in the art upon the reading and understanding of this specification and the annexed drawings. In addition, while a particular feature of the invention may have been disclosed with respect to only one of several implementations, such feature may be combined with one or more other features of the other implementations as may be desired and advantageous for any given or particular application.

The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. As used herein, the singular forms “a”, “an” and “the” are intended to include the plural forms as well, unless the context clearly indicates otherwise. Furthermore, to the extent that the terms “including”, “includes”, “having”, “has”, “with”, or variants thereof are used in either the detailed description and/or the claims, such terms are intended to be inclusive in a manner similar to the term “comprising.”

Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and will not be interpreted in an idealized or overly formal sense unless expressly so defined herein.

Claims

1-36. (canceled)

37. A controller for controlling a plurality of powered joints in a leg prosthesis, comprising: a processor; anda computer-readable medium, having stored thereon a plurality of instructions for causing the processor to perform the steps of: receiving, from a plurality of electromyogram (EMG) electrodes associated with a plurality of different muscle groups of a user, EMG signals for the user during a non-weight bearing mode of operation for the at least one powered joint;ascertaining an intent of the user based on a discriminant analysis of the EMG signals;generating a velocity reference for the at least one powered joint based on the EMG signals and the intent of the user; andtransmitting, to a motor unit for the at least one powered joint, a torque control signal for actuating the at least one powered joint, the torque control signal based at least on the velocity reference and the intent of the user.

38. The controller of claim 37, wherein the torque control signal is generated using a model based on the behavior of a spring and dashpot element.

39. The controller of claim 37, wherein the discriminant analysis is one of quadratic discriminant analysis of the EMG signals and linear discriminant analysis of the EMG signals.

40. The controller of claim 37, wherein the velocity reference is generated based on a principal component analysis of the EMG signals.

41. The controller of claim 37, wherein the torque control signal is generated based on an equilibrium point derived from the velocity reference.

42. The controller of claim 41, further comprising determining the equilibrium point based on the velocity reference and an initial angle of the at least one powered joint.

43. The controller of claim 37, wherein the at least one at least one powered joint comprises a plurality of powered joints, and further comprising alternating an active one of the plurality of powered joints when the EMG signals indicate a co-contraction.

44. A method performed in a controller for controlling at least one powered joint in a leg prosthesis, comprising: receiving, at the controller from a plurality of electromyogram (EMG) electrodes associated with a plurality of different muscle groups of a user, EMG signals for the user during a non-weight bearing mode of operation for the at least one powered joint;ascertaining an intent of the user based on a discriminant analysis of the EMG signals;generating, at the controller, a velocity reference for the at least one powered joint based on the EMG signals and the intent of the user; andtransmitting, from the controller to a motor unit for the at least one powered joint, a torque control signal for operating the motor unit and actuating the at least one powered joint, the torque control signal generated based at least on the velocity reference and the intent of the user.

45. The method of claim 44, wherein the torque control signal is generated using a model based on the behavior of a spring and dashpot element.

46. The method of claim 44, wherein the discriminant analysis is one of quadratic discriminant analysis of the EMG signals and linear discriminant analysis of the EMG signals.

47. The method of claim 44, wherein the velocity reference is generated based on a principal component analysis of the EMG signals.

48. The method of claim 44, wherein the torque control signal is generated based on an equilibrium point derived from the velocity reference.

49. The method of claim 44, wherein the torque control signal is generated based on an equilibrium point derived from the velocity reference and an initial angle of the at least one powered joint.

50. The method of claim 44, wherein the at least one at least one powered joint comprises a plurality of powered joints, and further comprising alternating an active one of the plurality of powered joints when the EMG signals indicate a co-contraction.

51. The method of claim 44, wherein ascertaining the intent of the user comprises classifying the EMG signals as corresponding to one of flexion or extension.