Open Access is an initiative that aims to make scientific research freely available to all. To date our community has made over 100 million downloads. It’s based on principles of collaboration, unobstructed discovery, and, most importantly, scientific progression. As PhD students, we found it difficult to access the research we needed, so we decided to create a new Open Access publisher that levels the playing field for scientists across the world. How? By making research easy to access, and puts the academic needs of the researchers before the business interests of publishers.
We are a community of more than 103,000 authors and editors from 3,291 institutions spanning 160 countries, including Nobel Prize winners and some of the world’s most-cited researchers. Publishing on IntechOpen allows authors to earn citations and find new collaborators, meaning more people see your work not only from your own field of study, but from other related fields too.
A critical point in the human–exoskeleton interfaces is the multivariable voluntary control of several joints independently. The lower limb exoskeleton ESROB, which helpes a patient to perform the sit-to-stand postural exercise, has been used for testing a new control based on electromyographic (EMG) signals and artifical neural networks (ANN). The approach is of “admittance control”, i.e. the joints of the exoskeleton are controlled in speed, instead of torque as usual, by mixing an automatic postural control loop (especially for the balance) with a voluntary action of the patient through EMG signals, measured on suitable muscles of the legs and of the trunk, processed by ANN. Mixing the automatic postural loop with the voluntary action by the patient helps during the training of ANN to exercise the different degrees of freedom of the exoskeleton and during the control to improve balance. This chapter describes the automatic postual control of ESROB as well as the experiments of training and of multivariable voluntary control by the patient. In particular, exploiting the separation offered by the algorithms, it is shown that the three degrees of freedom of the exoskeleton are controlled independently intermixing, the automatic control loop, through external sensors, and the voluntary control of the patient.
CRISP, Zhejiang University of Finance and Economics, Hangzhou, China
Massimo Mancin
Automation, Martoglio S.R.L.,Chiusa di San Michele, Turin, Italy
*Address all correspondence to: menga@polito.it
1. Introduction
The multivariable independent control of several degrees of freedom (DOFs) of an exoskeleton by the patient is a crtitical aspect of the human-machine interface, as it is easily deduced from the references contained in the next subsection.
In the present chapter, this problem is approached with a new control based on electromyographic (EMG) signals and artificial neural networks. The control is experimented on the lower limb exoskeleton ESROB for sit-to-stand exercises, realized in a previous research project.
To present the interface between the human and the exoskeleton, two aspects of biped robotics have to be considered: the DOFs of its kinematics (giving the dimensions of the two linked spaces of joints and Cartesian posture) and the balance. From the robotic point of view, the considered exoskeleton is relatively simple as it has only three DOFs, allowing the joint-legged movements of the two ankles, knees, and hips, corresponding to three postural variables in the sagittal plane of the Cartesian space, as show in Figure 1. That is to say the position of the projection on the ground of the center of gravity (COG) (and hence balance on the x-axis COGx), the height of the pelvis hz, and the angle of the trunk with respect to the vertical θtrunk.
Figure 1.
ESROB exoskeleton (a) in action, (b) its kinematics.
We presented a detailed analysis of the DOFs of a biped in these conditions in [1]. We also described the relationship that exists between COG, given by the angular position of the joints, and the center of pressure underfoot (COP) (also called “zero moment point” (ZMP)), given by the dynamics, and their implications on balance control in [2].
In the present approach, the joints of the exoskeleton are controlled in speed and position by combining an automatic postural control loop, closed from external sensors, with the voluntary action of the patient through “admittance control”, using the surface EMG signals measured on some of his muscles. The combination of the automatic loop and the voluntary action can be configured and from time to time changed, leaving some postural variables to the automatic control and the others to the patient’s voluntary action, which can be achieved through appropriate joints, or by blending the two controls on the same variables.
The “admittance control”, i.e. from torque to speed [3], for the man–machine interface is unusual, but it is necessary if it is to coexist with the automatic postural loop. It offers some advantages but, also, presents some drawbacks. Certaintly, it improves the automatic postural control on some DOFs, offering the advantage during the training to concentrate on single movements substaining automatically the patient on the others, but it limits the patient’s freedom. So the approach is useful for persons with strongh impaired balance and asks for their adaptation and physiological retraining.
The automatic postural control (COP position and consequent COGx), the height of the pelvis (substituted for simplicity by the angle of the knees), and angle of the trunk are obtained with a feedback from the measurements of the joint angles, from dynamometers under the feet, and by an accelerometer/gyroscope on the trunk [1].
For the voluntary action of the patient in “admittance control”, the applied torque is represented, approximately, by the EMG signals, and they are processed by an ANN. Two types of ANNs have been tested: NARX and LSTM. The output of the “admittance filter” implemented by the ANN controls in velocity the servomotors of the joints. Speed controls, unlike torque controls, offer postural stiffness. The patient is kept in a postural position. By pushing forward or backward on the exoskeleton, through the EMG signals generated from his muscles, he starts the motion which, once he has reached the desired final position, he has to stop, through a physiological feedback (to stop pushing).
The pairs of agonist and antagonist muscles chosen for each joint are: for the ankles the tibialis anterior (dorsiflexion) and the soleus (plantarflexion), for the knee the rectus femoris (knee extension) and biceps femoris (knee flexion), and for the hips the gluteus maximus (hip extension) and still the rectus femoris (hip flexion).
The approach followed in this chapter, using the functions offered by the ESROB control, is to record EMG signals, joint angular positions and velocities, while the patient performs a series of elementary exercises with voluntary motion of only one postural variable through one joint at a time (or at most two), while the other postural variables are controlled automatically by the postural loop. The recordings of all the elementary experiments are merged into a unique file to train a single multi-variable ANN capable to explain not only elementary motions but also complex ones involving multiple joints.
This chapter describes the techniques used, the training of the ANNs, the exercises performed by a healty person on the prototype, and the discussions of the obtained results.
1.1 State of the art and comparisons
A very recent survey [4] classifies the large body of literature for lower limb exoskeleton analysis and control. In this survey, two classifications are interesting for comparisons with the present approach: the “exoskeleton’s signal acquisition” and the “exoskeleton’s system modeling and control strategy”.
In the first classification, at the item of “prior signals volitional non invasive EMG” papers dealing with torque controls for 2 DOFs or walking speed and slope estimation for 1 DOF [5, 6, 7] are referenced. For “posterior signals” typically any type of of mechanical sensors can be integrated into the exoskeletons [8, 9].
The ESROB control merges “prior signals volitional non invasive EMG” with “posterior mechanical signals” using joint angle positions and load cells at the insole of the exoskeleton.
In the second classification, most of the papers are listed under the item “gait control” that does not apply here. Vice versa, similarities with the present approach exist with the papers at the item “mechanical impedance control” [10, 11, 12, 13], but admittance control has not been considered. Impedance is the relationship moving from velocity to torque, and torque is estimated and controlled. In this chapter, admittance control is used, i.e. the inverse relationship moving from torque to velocity, and velocity is the objective. In fact, admittance control is critical to guaratee postural balance.
Many papers propose estimating torque from EMG signals, e.g. [14]. Here, the EMG signals are used directly, as representative of torque (or the intention of the patient to move), in input of the admittance filter.
A certain number of papers are dealing with analysis and control of joint positions and velocities, using EMG signals and different kinds of ANN. Experimental examples of control hip [7] or knee [15], alone, were considered. All three joints, hips, knees, and ankles, as in this chapter, have been treated for analysis only in [16], but no practical experiments are presented.
A general discussion of admittance control for human–robot interaction, comparing this technique with impedance control, is available in [17]. The paper [18] approaches the control of a 4 DOFs (hips and knees of the two legs) of the exoskeleton. However, only the EMG signals of the knees are used through a central pattern generator to transform the network into one-dimension joint space admittance control. In fact, the paper claims there is too much complexity in N-dimension admittance controllers as well as incongruous movement of each joint. In the survey [9], two papers refer to admittance control [19, 20]. Both papers treat only the control of the ankle using EMG signals, the second, in particular, compares and shows the advantages of the EMG signals with respect to the direct measure of joint torque as input. Inside the EUROBENCH project, the authors [21] compare in the Exo-H3 exoskeleton an admittance control based on the trajectory error with a control triggered from EMG signals during walking on a treadmill.
To the knowledge of the authors, there have been no previous controls, apart from gait (where EMG signals are mostly used for classification or synchronization) or experiments of the voluntary multi-variable movements of all three joints of a leg (in the sagittal plane hip, knee, and ankle) when wearing a lower limb exoskeleton in actual operative postural positions during training. In fact, controlling all three joints is critical for balance. Another paper of voluntary multi-variable movements controls the torque only of hip and knee and carries out training without exoskeleton and in no operative postural position [22].
The experiments were performed here on a exoskeleton for sit-to-stand exercise because it was available, and it has only three DOFs, but, as is explained in the conclusions, it will be not difficult to free the two legs, separate EMG signals of both, and apply this technique to a complete exoskeleton.
The originality of the human–exosketeleton interface proposed in this chapter relies on its integration with an automatic postural control loop and in the linking of each joint with one corresponding DOF in the Cartesian space. Moreover in the integration, the DOFs between the automatic control loop and the voluntary joint actions can be configured in number, relationship, and intensity.
1.2 Organization of the chapter
The following Section 2 is technical: subsection 2.1 describes the control of ESROB, and subsection 2.2 presents the ANNs used for the experiments: NARX and LSTM.
The processing of recorded data and the types and sequences of exercises performed for training the ANNs are discussed in the sections 3 and 4.
Preliminarily, before the networks are implemented in the control, the performances of several experiments of training (optimization) only are evaluated. These experiments, in Section 5, are used for verifying the importance of input signals and for testing the hypothesis that networks trained by linking experiments with each single movement explain movements involving multiple joints.
Sections 6, 7, 8 contain the results of training the networks of NARX and LSTM and the results of using those networks for the voluntary control of independent 2 or 3 DOFs.
Finally, as an example of coordinate control between automatic posture and voluntary action, the sit-to-stand exercise is described in Section 9.
The conclusions, with the discussion of the results, are drawn in Section 10.
The MATLAB files of the last recorded datas and the films of the experiments are available to download at the end of the chapter, in the section titled “Additional materials”.
2.1 The control of ESROB: postural control and patient compliance
The control of ESROB integrates two aspects: balance and hapticity [1, 3, 23]. This is achieved with two feedback loops, one inside the other, as shown in Figure 2.
Figure 2.
Two control loops offering automatic balance and hapticity to the patient.
With x, ẋ indicate in the Cartesian space the vectors of coordinates or angles and velocities/angular speeds that uniquely define a posture and a motion of the joint patient–exoskeleton system (these have dimensions equal to the DOFs of the exoskeleton). The components of the vector are x=COGxθkneeθtrunkT, the projection of the COG onto the ground, the height of the pelvis (for simplicity θknee), and the angle of the trunk. With θ=θankleθkneeθhipT, θ̇, in joint space, indicate the angles and angular velocities of ankles, knees, and hips. A classical linear relationship exists between the two velocity vectors ẋ=Jθ⋅θ̇, where the square matrix J, a nonlinear function of θ, is called Jacobian. For any consistent postural position, with the exception of a few singular points where an extended Jacobian has to be used, as described in [24], J is invertible, so that θ̇=Jθ−1⋅ẋ, through which, from a desired postural motion, the corresponding motion of the angles of the kinematic chain is obtained.
Therefore, looking at Figure 2, the outer loop, operating in the Cartesian postural space, automatically controls posture (and therefore also balance), and the inner loop, which offers “compliance”, acts directly on the joints of the exoskeleton with patient-operated admittance control via surface EMG signals measured on a number of muscles in the legs and the trunk.
The block “Preview” contains the desidered trajectory (positions and velocities) of the postural variables (eventually velocity is zero if a variable is maintained constant). The block “Ks” contains the gains and the compensation filters of the postural loop. “Gs” includes the compensation filters of the joint velocity loops.
On the other hand, the two functions, maintaining a posture and offering compliance, clearly compete during a rehabilitation cycle when the patient is impaired. In fact, the exoskeleton must be able to offer stiff postural support in the initial stages and gradually becomes more “compliant”, i.e. enslaved to the patient, as he gradually regains his motorial functionalities. At the end of the rehabilitation cycle, the external postural loop simply has to monitor balance and intervene, through the block “Inhibitor” (admittance is set to zero), only when the limits are reached.
To achieve this result, the interactions between automatic postural control and voluntary joint control by the patient can be configured at two levels:
separating in the postural space two subspaces in which automatic control on the one hand and voluntary control, on the other, act independently, and contemporaneously define the appropiate subspace in the joint space operated by the patient;
moreover, when the two controls coexist in the same subspace, merging, through a parameter indicated with β (0≦β≦1), the postural loop and the cooperative action of the patient.
To ensure a posture and balance so that the biped does not fall down, or to perform a postural task, a “whole body coordination” (WBC) [25] is postulated in biped robotics. In the realm of the WBC, Choi et al. [26] introduced the kinematic resolution of the Jacobian of the COG with embedded movement. In this way, all DOFs are controlled.
Here, an original reinterpretation of the WBC is introduced which is called “tutoring-cooperation-coordination”: where some postural variables through appropriate joints are under the patient’s control, and the complementary ones are controlled separately by the automatic postural loop.
Tutoring means that the exercise is completely performed by the automatic postural loop, through a program contained in the block Preview of Figure 2, with a passive patient.
Coordination is when the patient has complete control, through an equal number of joints, of one or more components of the postural task of the exercise, without interfering with the complementary components controlled by the outer loop, and therefore, he must perform a coordinated movement with the postural loop to generate a meaningful exercise.
Between the two extremes, Cooperation means that on some components of the postural task the two actions are contemporaneously present and overlap. The patient has partial control over these components, with intensity modulated by the coefficient β, and must cooperate with the block Preview.
To obtain the desired decoupling of postural and joint variables in the interaction, partially modify the block diagram of Figure 2, replacing the dotted part 1 with 3a of Figure 3 and detailing the admittance control 2 with the artificial neural network 3b, where, in Figure 3a, ẋdesired are the desired postural movements, θ̇p is the patient’s contribution, as an output of the admittance control, θ̇ref is the reference, input of the motor speed servo mechanisms, Ru,Nu are diagonal selection matrices with, one in the range and in the null-space, respectively, otherwise zero, of the postural tasks ẋ automatically controlled in the Cartesian space. Rp,Np are analogous matrices referring to the patient contributions θ̇p in the joint space. For an example, if the automatic postural loop maintains the position of the COP (hence controls the balance), and the patient controls the height of the pelvis and the angle of the trunk from the knees and the hips, with the order of variables in the Cartesian space and in the joint space, specified at the beginning of this section, the matrices are:
Figure 3.
Tutoring-cooperation-coordination - (a) is the modification to the control to achieve indendency of the automatic posture and the patient action, (b) the admittance filter is implemented by a ANN, and (c) the schematical representation of the interaction of the two controls.
Ru=100000000Nu=000010001Rp=000010001,E1
F=Ru⋅J is an algebraic matrix, and 0⩽β⩽1 a coefficient that provides the intensity of the patient’s contribution (0 = no contribution-tutoring, 1 = total contribution-coordination, intermediate values = cooperation). The value of β, obviously, has consequences on the admittance felt by the patient (β=0 completely stiff system, and β=1 system fully “compliant”). The proof is in Appendix A.
With an appropriate choice of Ru,Rp and β, some of the components in the postural space are automatically controlled by the outer loop, and the others are controlled independently with coordination or in cooperation, by the patient through his EMG signals. As an extreme case with Nu=Rp=I and β=1, the entire exoskeleton is controlled by the patient. An interesting case is when Ru and Rp are chosen, as in Eq. (1), so that the patient controls all the components to perform a postural exercise through all joints but the ankles. While balance, on the other hand, is automatically guaranteed by the postural loop. Actually, this technique will be used at the end of the chapter for the “sit-to-stand” exercise, leaving the patient to control the height of the pelvis and the trunk angle, through the knees and hips and the postural loop to control balance (COGx−COP).
The interaction between the two players (the automatic system and the patient) can be summarized in Figure 3c.
Finally, the patient’s contribution θ̇p is obtained by finding muscle synergies [27, 28] from the EMG signals taken from appropriately chosen muscles from the legs and trunk (not applied in these experiments) and training neural networks, to be used for real-time control, as shown in Figure 3b.
2.2 Adopted neural networks
Neural networks implement the multivariale admittance control filter of the patient–exoskeleton interface. The inputs are the EMG signals of the patient’s muscles involved in the movement (indicators of the intention, speed, and intensity of performing joint movements) and possibly the joint angle position values and the outputs the angular velocities used as references to the speed servomotors of the joints of the exoskeleton. The training of these networks is achieved by joining together the data of a series of partial exercises (sub-experiments, at the limit voluntary movements of single joints) performed by the patient, recording EMG signals, angular positions, and velocities of the joints. The speed references, output of the filter, initiate the movement of the exoskeleton joints, and it is then the patient’s task (always through the EMG signals) with a physiological reaction to make them to stop once they have reached the desired position. Finally, as it happens in the training of neural networks, a part of these signals is used for the actual training, a part for the validation, and a third part for the final test (not used in these experiments). This is to avoid over-parameterizing the network or biasing it too much towards the specific data used for training.
The two types of networks tested in the experiments were the NARX and the LSTM. As MATLAB toolkits are used, all the information regarding these networks can be found there.
2.2.1 The NARX networks
NARX are nonlinear autoregressive neural networks with external moving average inputs. As parameters, the number of delays of the samples of the autoregressive part and of the moving average part, and the number of hidden layers of the network are assigned. The number of delay samples in the network allows the dynamics of the system to be taken into account, and the number of hidden layers considers the complexity of the phenomenon. In these networks, it is possible to assign a different number of delays to different groups of inputs (as in Figure 4). This is particularly useful in our case to treat EMG signals and joint angle positions differently, as described in the following sections.
Figure 4.
Layout of the NARX network used in the experiments.
The optimization was performed using the Leveberg–Marquardt algorithm with hyperparameters given in Table 1.
Minimum gradient
1e-07
Maximum validation checks
16
Mu
0.001
Mu decrease ratio
0.1
Mu increase ratio
10
Maximum mu
10,000,000,000
Table 1.
Hyperparameters for training NARX networks.
2.2.2 The LSTM networks
Long short-term memory (LSTM) is a recurring artificial neural network used in the field of “deep learning”. Unlike “feed-forward” standard neural networks, LSTMs, as well as NARXs, have a feedback connection, and the outputs are brought back to the input, as it occurs in autoregressive filters. The design parameters are the number of inputs, the number of outputs, and the number of hidden layers. The dynamics of the phenomenon in these networks is captured by the hidden layers, in a way not completely transparent, and it is not possible to treat in input EMG signals and joint angle positions, as in NARX networks, differently (Figure 5).
Figure 5.
Example of the layout of a LSTM network.
The optimization was performed using the Adam (adaptive moment estimation) optimizer, including learning rate information, L2 regularization factor, and mini-batch size algorithm with hyperparameters given in Table 2.
Each experiment consists in recording the EMG signals, position, and speed of the joint angles in a series of elementary exercises performed by the patient wearing the exoskeleton as described in the following sections; in training a neural network from the data; in implementing it in the control code of the exoskeleton and in testing the control on the same patient who provided the training data. It should not be overlooked that the patient–exoskeleton interface is then personalized to the specific individual and his characteristics.
The degree of success is not only evaluated from the performance obtained in the trainining but especially is subjectively evaluated qualitatively by the patient on the ease of performing both the exercises for which the net has been trained and other more complex exercises. As an example, by introducing in input a number of delays of the joint angle samples helps, in NARX networks, to improve the performance in training, but the patient has more difficulty to voluntary control the movements. This is because the angles in input correlate with the velocities in output, and this reduces the predictive power of EMG signals. For this reason only current angle positions are adopted in these networks. The same phenomenon does not occur for LSMT networks, where the two groups of inputs, EMG signals, and angle positions cannot be treated differently, and this aspect is under further investigation.
Usually recordings have been made in training experiments on moving voluntarily single joints and, even if the patient has not complete control of the exoskeleton, on pairs of joints in order to improve training or simply for “validation”.
3.1 Processing of EMG signals
The motion was on the sagittal plane, only, with joint legs. For this, two approaches have been tested with similar results: averaging the corresponding EMG signals on the two legs; measuring the signals on just one leg.
The measures, as all the real-time control, were done with a sampling frequency of 1 KHz. The EMG signals were pre-processed by a band pass filter (100–150 Hz), rectified, and further filtered with a low pass filter of 100 rad/s.
However, the recording of pre-processed EMG signals, positions, and angular velocities was performed in no strict real time with an average sampling period of 0.02 sec. The absence of real time needs a resampling at exactly 0.02 sec of the signals before applying them to the network training algorithm. The control of the exoskeleton is obviously in real time at 1 ms. The filters obtained, based on 0.02 sec samples, therefore require a moving average operation on 20 input and output samples.
3.2 Normalization of the inputs
All input signals are normalized. Two normalization techniques have been used with very similar results: from each signal its average value is subtracted, and then, the zero mean signal is divided by its standard deviation, or simply, without taking it to zero average value, it is divided by its maximum amplitude (maximum minus minimum values).
3.3 Elemental esercies
The functionalities of the exoskeleton control, described in Section 2.1, were used, which allow automatic postural controls and voluntary actions of the patient to be merged together (“Coordination”, i.e. β=1). During training, elementary voluntary exercises were performed one joint at a time or, exceptionally, pairs of joints, while the remaining DOFs were maintained automatically to a constant value through the postural loop. Indeed, without an already trained network, it is not possible to move all three joints voluntarily, independently, in a coordinated way.
The following experiments were performed:
voluntary movements of the COGx through the ankles by automatically maintaining the height of the pelvis (angle of the knees) and the angle of the trunk;
voluntary movements of the knees (height of the pelvis) automatically holding COP (balance) and trunk angle;
voluntary movements of the angle of the trunk through the hips, automatically maintaining the COP and the angle of the knees;
voluntary movements of COGx and height of the pelvis through ankles and knees with automatic holding the angle of the trunk;
voluntary movements of COGx and angle of the trunk through ankles and hips maintaining with automatic control the height of the pelvis;
voluntary movements of the height of the pelvis and angle of the trunk through knees and hips with automatic control of the COP.
Remember that the postural position is guaranteed by the integration of the two controls, and some DOFs are rigidly controlled by the postual loop. The patient has to concentrate his attention on the only (or only two) joint/s (in which the admittance is high) and on the voluntary movements that he can perform. Unlike other works on training for voluntary controls using EMG signals, the exercise is performed here in functionally operative postural positions (i.e. under the effect of the gravity) and in the presence (i.e. with the constraint) of the exoskeleton and its control.
Regardless of voluntarily controlled muscles the entire six-elements vector of EMG signals, all angles, and all angular velocities are recorded.
3.3.1 How EMG signals are processed in training
Since the joints are speed-controlled, the patient is constrained to the postural position of the exoskeleton. In training, the postural variables under test are released from the postural loop, and the joints under patient control are independently actuated through the two EMG signals of the corresponding pair of muscles, agonist, and antagonist. Since a trained filter does not exist yet, the control action is obtained by experimentally implementing for each joint a simple admittance filter of the type:
velreft=fAgot⋅kago−fAntat⋅kanta+h⋅K,E2
where the reference velref is applied to the velocity control of the joint, f is a function with unitary gain and a central dead band, Ago and Anta are the two pre-processed EMG signals, agonist and antagonist, related to the considered joint, kago,kanta,K are gains, and h is an offset.
The patient tries to perform the exercise, with maximum naturalness and minimum effort: the dead zone of the function f and h are regulated to allow to remain stationary at rest, the gains kago,kanta to equalize the effort in both directions, and K to minimize it.
In an attempt to move, the patient excites the corresponding muscles by forcing on the exoskeleton link, and with a reduced effort he believe to perform the movement of the joint and the postural coordinate under his control, and in reality the servomotor of the joint has intervened, instead. Simultaneously, the postural loop controls the remaining postural coordinates.
Slow and fast movements are performed, with variable amplitude, interspersed with periods of rest. The EMG signals related to all joints, both voluntarily and automatically moved, are recorded, together with joint velocities and angular positions.
3.3.2 Multiple movements of joints
As already been noted in the ESROB project, the technique described in Section 3.3.1, where the joints are independently controlled by pairs of EMG signals from corresponding antagonistic muscles, the patient allows to move one joint at a time, and very approximately two joints, while when all three joints are controlled at the same time there are incongruous movements. So in the training, in addition to the experiments on single joints, some movements on two joints have been added.
The experiments were made on an healthy person. In fact, one of the authors offered himself. Twenty-one training sessions were carried out between November 2021 and October 2022, with a total of 31 admittance filters being only identifed or tested in the exoskeleton control. The initial experiments were used to choose the acquisition operating modes, the suitable parameters for the filters and to compare the two types of neural networks. The last acquisitions of 27/06/22, 21/07/22, and 02/08/22 generated the filters for the final experiments, illustrated in this chapter.
A typical example of an experiment (21/07/22) consists of repeating three times similar groups of voluntary elemental movements (ankles only, knees only, and hips only) and one repetition of three double movements (ankles–knees, ankles–hips, and knees–hips). Angles, speeds, and pre-processed normalized EMG signals of the whole experiment are presented in Figure 6, with details of the joint velocities and EMG signals in Figure 7. These belong to the sub-experiment where the patient is voluntary moving the ankles and knees and the automatic postural control maintains the angle of the trunk constant.
Figure 6.
A typical example of data acquisition for training, where each sub-experiment is evident—the experiment of 21/07/22. (a) the joint angle positions, (b) the joint angle velocities, and (c) the normalized EMG signals.
Figure 7.
Details of the recording of the experiment of 21/07/22—voluntary motion of the ankle and knee—(a) joint angle velocities, (b) EMG signals.
The inputs of the network are the six EMG signals and the three joint angular positions, and the outputs are their angular velocities. The input EMG signals (true predictors of the process) and the output angular velocities require a number of delay samples, necessary to capture the dynamics of the phenomenon. The joint angles, on the other hand, only serve to qualify the instantaneous position where the signals were recorded and must not create a correlation with the output angular velocities, so only the current sample and not their delayed samples have to be used. In NARX networks, assigning a number of different delay samples to groups of different inputs is possible. In LSTM networks to indicate the number of delay samples for each input is not possible, as the past is stored, without distinction, for all the inputs in the hidden layers of the network.
The entire registration is divided into two sections: one part that is used for training and the other part for validation to allow the over-parameterization of the network1 to be verified. Furthermore, to avoid the network being influenced too much by the training data, it is convenient in the training process to keep the two cost functionals (the mean squared error between the output data and their estimates) of training and validation separate and to stop the optimization when the minimum of the validation functional is reached. Both optimization frameworks of NARX and LSTM incorporate these features. In fact, in all the figures, where the process of optimization is presented, training and validation performances are shown separate, and the process is stopped when the validation performance reaches the minimum.
Before using the neural networks in the control, some training and validation experiments were carried out with the aim of verifying the initial hypotheses. The two types of networks have different characteristics and have been analyzed separately.
5.1 NARX networks
When initial experiments were completed, 5 input and output delays and 10 hidden layers were chosen for the networks used (Figure 4), whether or not input joint angles were present. Two aspects were examined: the importance of joint angle signals in input and the possibility of using elementary motion experiments to explain complex motions. The indices used are the quadratic error mean (MSE) of the validation performance index and the percentage to which the speed prediction explains the speed used for training (one minus the square root of the mean squared error over the square root of the mean square of the signal).
5.1.1 The presence of joint angle signals at the input
The experiment from 27/06/2022 was used for this analysis. Data were obtained by merging three repetitions of similar groups of voluntary elemental movements (ankles only, knees only, hips only) and three repetitions of two groups of double movements (ankles–knees, knees–hips). Training was using 75% of all sub-experiments, while the 25% terminal part of each one was used for validation.
Three tests were performed with different settings of the angle positions in input. The results were in Table 3.
MES of validation
Explanation
Samples of joint positions in input
3.0
97%
number of delayed samples the same as EMG signals
11.8
86%
only present time sample
17.0
78%
no joint positions in input
Table 3.
Performances with different settings of joint angles position in input using NARX networks.
The presence in input of the joint angles, obviously, improves the fitting. However, using past samples helps to correlate inputs angle positions and output angular velocities, and in these conditions, the voluntary control of the exoskeleton by the patient has been found to be more difficult. Therefore, the solution of using the present sample only has been adopted in the following.
5.1.2 Simple and complex movements
The analysis provided here refers to the acquisition experiment of 21/07/22 in Figure 6. The typical trend of the training performance index in the various optimization phases (epochs) is given in Figure 8. Even if the performance index of training continues to decrease during the iterations, the process is stopped when the validation performance starts to increase. Afterwards the network explains the noise present in the part of the signal used for training, but not the essential behavior of the phenomenon.
Figure 8.
Typical behaviors of the MSE for training and validation data in a NARX network—in this example all three speeds are estimated using all sub-experiments, splitting 75%, 25% between training and validation.
The results obtained by performing training on parts or on the entirety of the samples are detailed below.
In the first five experiments training and validation periods for all sub-experiments were 75%, 25%. In the final experiment, training was done on the whole single movements and validation on the whole double movements. Performances of the three networks with data limited to the three sub-experiments with voluntary movement of only one joint, and estimation of the corresponding joint velocity is contained in the Table 4.
MSE of validation
Explanation
Joint velocity
0.3
98%
ankle
1.8
99%
knee
3.6
92%
hip
Table 4.
Performances estimating each single joint velocity using NARX networks.
Performances of a single network explaining the 3 voluntary movements, with different conditions, are contained in the Table 5.
MSE of validation
Explanation
Condition of the experiment
7.3
83%
excluding the sub-experiments with the double movements
13.2
77%
including the sub-experiments with the double movements in training and validation, therefore, the double movements are also included in the training
31.7
70%
training on the single voluntary movement sub-experiments and validation on the double movement sub-experiments, and therefore, the double movements are only used for validation and not for training
Table 5.
Performances estimating a unique network for all veloicities with different settings using NARX networks.
As can be seen, performance degrades when moving from single angle voluntary movement to multiple single angle and double angle movements. However, the last result is interesting, because it highlights how training on single voluntary movements is, to a certain extent, able to interpret even complex movements.
5.1.3 Concluding remarks
Based on the experience in testing the networks, as the input presence of the joint angles with their delays reduced the predictive ability of the EMG signals, angles in input without delay samples were adopted. Simple and double movements were used with the periods of 75%, 25% of each sub-experiment, between training and validation.
5.2 LSTM networks
LSTM networks cannot handle differently EMG signals and joint angles in input. Therefore, experiments were performed with and without input angles and with different values of the number of hidden layers.
A number of hidden layers equal to or greater than 100 is needed to have an acceptable prediction of the outputs and reasonable bandwidth of the filters, but, for data size issues, with 200 hidden layers the exoskeleton real-time processor is not able to run the control program. Therefore, 100 hidden layers have been chosen.
Apparently, the presence or absence of input angles did not provide appreciable differences in the performance of the filters. In particular, different to NARX networks, the presence of input angles in real-time control applications did not reduce the predictive power of EMG signals. For this reason, the angles at the input were left in the final real-time experiments, with periods of 75% and 25% of each sub-interval between training and validation.
Figure 9 and the results of Table 6 refer to the identification of the three velocities with a unique network, from the recordings of 27/06/22, using all single and double joints voluntary movement sub-experiments, with a validation on the final parts of each one.
Figure 9.
Trend of the performance index during optimization of a LSTM network with 100 hidden layers—in the presence (a), and without (b) joint position angles in input.
Inputs
Ankles
Knees
Hips
with angles
0.79%
0.89%
0.87%
without angles
0.78%
0.88%
0.86%
Table 6.
Percentage of explanation of output velocities in training using LSTM networks.
6. The identification of networks with three degrees of freedom
The identification results of the two networks used in the final 3 DOF experiments are shown. Both are based on 21/07/22 data. Training is carried out using all data, single and double voluntary movements, excluding the last parts of each sub-experiment used for validation, with the usual 75%, 25% split between training and validation. The accuracy of the estimates on each single motion sub-experiment is very good and scarcely significant. Therefore, the following figures have been chosen to show the results of the estimates of the outputs for training and validation data only for the double movements.
6.1 NARX
The NARX network has in input 5 EMG signal delay samples and the angular positions of the joints to the present sample only, in output 5 velocity delays, and 10 hidden layers.
The training performance is in Figure 10. The following Figure 11 shows the estimates of the output speeds.
Figure 10.
Trend of MSE in training and validation data, for the identification of a NARX network predicting three velocities.
Figure 11.
Velocity estimates in double movements during training by the NARX network used for the final experiments—(a) ankle–knee training, (b) ankle–knee validation, (c) ankle–hip training, (d) ankle–hip validation, (e) knee–hip training, (f) knee–hip validation.
6.2 LSTM
The network has 9 inputs (the 6 EMG signals and the three angle positions) and 100 hidden layers. The training performance is in Figure 12.
Figure 12.
Training performance.
The following Figure 13 shows the estimates of the output speeds in double movements. For single movements, they are very good and not significant.
Figure 13.
Velocity estimates in double movements during training by the LSTM network used for the final experiments—(a) ankle–knee training, (b) ankle–knee validation, (c) ankle–hip training, (d) ankle–hip validation, (e) knee–hip training, (f) knee–hip validation.
A non-optimal filter bandwidth is noted, and an increase in the number of hidden layers beyond 100 would certainly have improved the response.
The results of three exercises are presented in this section. The first two exercises, with 2 DOFs, were obtained with 2 and 3 DOFs preliminary filters estimating only two velocities, or using only two velocities and the third with the NARX filter trained in the previous Section 6. Three videos are available.
7.1 Exercises with two degrees of freedom
7.1.1 Ankle and knee
The two DOFs model, i.e. only two joint speeds are estimated and controlled, is one of the first used and was built 01/03/22 with data dated 28/02/22. The exercise was performed 06/10/22 and the video is available at \videos\NARX ankle_knee_2022_10_061.mp4.
Voluntary movements were performed to move repetitively the height of the pelvis and simultaneously move the center of pressure underfoot from heel to toe. The trunk is kept vertical by postural control.
To highlight the patient’s ability to control the ankles and knees, an attempt was made in the movements to make the pelvis follow circular trajectories in the Cartesian space, as shown in Figure 14.
Figure 14.
Trajectories of the pelvis in the x-z sagittal plane with control by the patient of ankle and knee using a NARX filter, with the objective to obtain a circular path—(a) first experiment and (b) second experiment.
Figure 15 shows the three angles of the joints and the trunk angle, which is maintained within one degree by the postural control. Figure 16 shows the changes in pelvis height and the center of pressure underfoot.
Figure 15.
Angles of interest during ankle and knee movements by the patient in the previous experiment—(a) the three angles of the joints, (b) the trunk angle with respect to the vertical.
Figure 16.
Positions of interest during ankle and knee movements by the patient in the previous experiment—(a) vertical coordinate of the pelvis, (b) COP under the feet on the x-axis.
7.1.2 Knee and hip
Only 2 DOFs, of a 3 DOFs model, built from data dated 13/07/22, with experiments dated 06/10/22, was used for the patient control. The video is available at \videos\NARX knee_hip_2022_10_06.mp4.
A movement was performed to change repetitively the angle of the trunk and simultaneously move the height of the pelvis. The center of pressure under the feet is maintained by postural control (Figure 17).
Figure 17.
Variables of interest during hip and knee movements by the patient with a NARX filter—(a) behaviors of the three angles of the joints, (b) trunk angle, (c) vertical coordinate of the pelvis, and (d) position of the COP under the feet on x axis, maintained automatically from the postural control.
7.2 Exercises with three degrees of freedom
The model used data recorded on 21/07/22, built the same day, and tested on 06/10/22. All three joints are under the control of the patient who, while maintaining balance, simultaneously changes the angle of the trunk and the height of the pelvis. The video is available at \video\NARX 3GradiLiberta_2022_08_02_16_07.mp4.
Figure 18 shows the movements of the angle of the trunk and the height of the pelvis, keeping the COP, reasonably, inside the soles of the shoes.
Figure 18.
Variables of interest during the movements of the three joints by the patient with a NARX filter—(a) behaviors of the three angles of the joints, (b) trunk angle, (c) vertical coordinate of the pelvis, and (d) COP under the feet on the - axis.
The LSTM networks were tested directly with 3 DOFs.
8.1 Exercises with three degrees of freedom
Model was built 08/10/22 on data from 21/07/22, run 11/10/22 with a video from the same date at \video\LSTM LSTM_9_3_2022_10_11.mp4
The characteristic variables of the exercise are given by the following Figure 19.
Figure 19.
Characteristic variables of the movement of 3 DOFs by the patient with a LSTM filter—(a) angles of the joints, (b) trunk angle, (c) x and z coordinates of the pelvis, and (d) COP under the feet.
For the sit-to-stand exercise, the 3 DOFs NARX filter trained on 02/08/22 with data of 02/08/2022 was used. Only 2 DOFs were implemented in real time, leaving knee and hip control to the patient and performing an automatic control of the COP with the postural loop. Two videos are available at \videos\NARX SitToStand1_2022-08-02_16-01-35.mp4 and SitToStand3_2022-08-02-16-07-03.mp4.
Figures 20 and 21 show the Cartesian and joint variables during the exercise, respectively.
Figure 20.
Variables of posture on the Cartesian space during sit-to-stand with kees and hips controlled by the patient using an NARX network, and the postural loop guaranteeing the balance—(a) behavior of the COP under the feet, (b) trunk angle, (c) position of the pelvis on the x axis, and (d) position of the pelvis on the z-axis.
Figure 21.
Joint angles during sit-to-stand with kees and hips controlled by the patient using an NARX network, and the postual loop guarateeing the balance—(a) ankle angle, (b) knee angle, and (c) hip angle.
The exercise sit-to-stand was performed in three phases.
phase 1 Complete postural control over the three DOFs. The patient was brought seated with trunk upright. The COG is obviously set back from the feet support.
phase 2 In the previous seated position, at instant 20,175 sec the patient, with only control of the hips, was asked to bend his trunk forward, and contemporaneously the postural control brought the COG to the center of his feet.
phase 3 At instant 20,182 sec the 2 DOFs voluntary control was activated. As well as hips, the patient had control of his knees, while the postural control maintained the COG at the center of the feet. In these conditions, the patient was asked to stand up and to move the trunk upright.
The experiments to control an exoseleton for sit-to-stand exercises involving voluntary motion of multiple joints are presented. In order to improve the balance for strongly impaired persons, control merges voluntary action by the patient with EMG signals processed by ANN, with an automatic postural control from external sensors. In spite of the wide ranging literature of using EMG signals processed by ANN in exoskeletons, the approach is original on three aspects:
using for compliance “admittance control” and controlling joint speed, instead of “impedance control” and torque, as usual;
mixing voluntary action and automatic postural control, independently on different DOF of the kinematics;
attempting to voluntary control multiple joints, the training of ANN networks is performed merging all single joint movements, on the identical postural position and in the presence of the exoskeleton for the future real use.
Two recurrent neural networks, NARX and LSTM, have been tested for the voluntary control of the exoskeleton in the postural exercises presented in this chapter. For easy to perform the exercise NARX nets took advantage of using the current joint angle positions in inputs with no past samples. On the other side, LSTM showed not significant differences in the presence or absence of joint angles in input.
The two nets behaved almost equally well. However, it is noted that LSTM nets require a much greater number of hidden layers and consequently memory footprint than NARX ones. This should not be a problem with future processors.
The initial intentions of the research were to design the exoskeleton and its control technique for rehabilitation, i.e. leading a patient with trauma to recover to the normal state and then eliminating the exoskeleton.
However, it has been verified that the presence of the exoskeleton, above all with speed control on the joints and consequent stiffness in the posture, while it potentially improves the balance with the postural control loop, permanently modifies the feeling perceived by the patient and his physiological behavior. Subsequently the study has been addressed to the development of a human–computer interface to allow weak or impaired people to move maintaining the balance wearing the exoskeleton permanently. This requires not only network training, but also leads to user adaptation.
The results are encouraging, but not completely satisfactory. In particular, the use of surface EMG signals has shown some limitations:
the noise on the EMG signals;
the difficulty of getting signals from deep muscles. For example, the movement in the sagittal plane of the hips is controlled along with the gluteus by the ilio-psoas group which are deep muscles. These have been replaced by the rectus femoris which, however, also controls the knee.
In order to overcome these limits, deep EMG signals are needed with the use of needle electrodes [29], clearly more intrusive and not available to current experimenters.
These first experiments were conducted on a perfectly able person. The next experiments on this exoskeleton will consider people with different kinds of illnesses, in particular with asymmetry on the two legs, and signal acquisition on only the most able one. Moreover, it is expected to digitally clean the input data in the presence of spasticy, that is frequently present in the case of accidents. As a further step, a lower limb exoskeleton with all degrees of freedom (ten) to allow walking in a rectilinear path will be created. It should not be more complicated, using the same EMG signals as those used here. It is necessary just to keep the EMG signals of the two legs separated.
Acknowledgments
The exoskeleton ESROB has been developed by the support of the project ESROB (PAR FSC 2007/13) of the Italian region Piedmont. This research was supported by the feasibility study ESROB-2 (F.E.S.R. 2014/2020) of the same region.
Thanks are given to Martoglio S.R.L. for offering the technical support for conducting the experiments.
Additional materials
The video recordings and Matlab data files referenced in this chapter can be downloaded from this link: https://bit.ly/3AJ2g9Z
Appendix A: proof of the separation on the Cartesian space between automatic and voluntary controls
Let move in the block diagram of Figure 3a the signal θ̇1p bringing it before the block J−1. According to the block diagram algebra, the contribution of θ̇p is, therefore, I−Ru⋅J⋅Rp⋅β⋅θ̇p, noting that I−Ru=Nu. In conclusion, the signal at the input of the block J−1 becomes Ru+Nu⋅1−β⋅ẋ+Nu⋅J⋅Rp⋅β⋅θ̇p. This expression proves that with β=0 there is no patient contribution in the Cartesian postural space. With β=1, automatic postural control ẋ only acts in the range of Ru and that of the patient θ̇p in its null−spaceNu, i.e. they are completely decoupled.
Obviously, Rp, that defines the patient’s controlled joints, cannot be chosen arbitrarily. To have correct posture when β>0, the choice of patient’s controlled joints, in the range di Rp, must satisfy the condition rangeNu⊂rangeJ⋅Rp.
With intermediate values 0<β<1, the postural variables in the range of Ru remain under the exclusive control of the automatic postural loop, but those in the range of Nu are jointly controlled by the automatic postural loop and the voluntary action of the patient, with the value of β that blends the two contributions.
References
1.Menga G, Ghirardi M. Control of the sit-to-stand transfer of a biped robotic device for postural rehabilitation. Robotics. 2019;8(4):91. Special Issue Advances in Italian Robotics. DOI: 10.3390/robotics8040091
2.Menga G, Ghirardi M. Estimation and closed-loop control of COG/ZMP in biped devices blending CoP measures and kinematic information. Robotics. 2019;8(4):89
3.Menga G, Ghirardi M. Lower limb exoskeleton for rehabilitation with improved postural equilibrium. Robotics. 2018;7(2):28. ISSN 2218-6581
4.Sun Y, Tang Y, Zheng J, Dong D, Chen X, Bai L. From sensing to control of lower limb exoskeleton: A systematic review. Annual Reviews in Control. 2022;53:83-96
5.He H, Kiguchi K. A study on EMG-based control of exoskeleton robots for human lower-limb motion assist. In: 2007 6th International Special Topic Conference on Information Technology Applications in Biomedicine. Tokyo, Japan; 2007. pp. 292-295. DOI: 10.1109/ITAB.2007.4407405
6.Liu H, Gui K, Zhang D. A practical & adaptive method to achieve emgbased torque estimation for a robotic exoskeleton. IEEE/ASME Transactions on Mechatronics. 2019;24(2):483-494
7.Kang I, Kunapuli P, Hsu H, Young AJ. Electromyography (EMG) signal contributions in speed and slope estimation using robotic exoskeletons. In: 2019 IEEE 16th International Conference on Rehabilitation Robotics (ICORR). Toronto, ON, Canada; 2019. pp. 548-553. DOI: 10.1109/ICORR.2019.8779433
8.Hussain S, Kapsalyamov A, Jamwal PK, Ghayesh M. State of the art lower limb robotic exoskeletons for elderly assistance. IEEE Access. 2019;7:95075-95086
9.Ijspeert A, Baud R, Manzoori AR, Bouri M. Review of control strategies for lower-limb exoskeletons to assist gait. Journal of Neuroengineering and Rehabilitation. 2021;18(1):119
10.Huo W, Mohammed S, Amirat Y. Impedance reduction control of a knee joint human-exoskeleton system. IEEE Transactions on Control Systems Technology. 2019;27(6):2541-2556
11.Huo Y, Li X, Zhang X, Sun D. Intention-driven variable impedance control for physical human-robot interaction. In: 2021 IEEE/ASME International Conference on Advanced Intelligent Mechatronics (AIM). IEEE Press; 2021. pp. 1220-1225. DOI: 10.1109/AIM46487.2021.9517438
12.Chen L, Wang C, Song X, Wang J, Zhang T, Li X. Dynamic trajectory adjustment of lower limb exoskeleton in swing phase based on impedance control strategy. Proceedings of the Institution of Mechanical Engineers, Part I: Journal of Systems and Control Engineering. 2020;234(10):1120-1132
13.dos Santos WM, Siqueira AAG. Optimal impedance via model predictive control for robot–aided rehabilitation. Control Engineering Practice. 2019;93:104177
14.Wu Q, Xu D, Zhu Y. Development of a semg-based joint torque estimation strategy using hill-type muscle model and neural network. Journal of Medical and Biological Engineering. 2021;41:34-44. DOI: 10.1007/s40846-020-00539-2
15.Gallagher S, Zabala M, Coker J, Chen H, Jr MC, Schall. EMG and joint angle-based machine learning to predict future joint angles at the knee. Sensors. 2021;21:3622. DOI: 10.3390/s21113622
16.Chen J, Zhang X, Cheng Y, Xi N. Surface EMG based continuous estimation of human lower limb joint angles by using deep belief networks. Biomedical Signal Processing and Control. 2018;40:335-342
17.van der Kooij H, Keemink AQL, Stienen AHA. Admittance control for physical human–robot interaction. The International Journal of Robotics Research. 2018;37(11):1421-1444. DOI: 10.1177/0278364918768950
18.Gui K, Liu H, Zhang D. A generalized framework to achieve coordinated admittance control for multi-joint lower limb robotic exoskeleton. In: 2017 International Conference on Rehabilitation Robotics (ICORR), London, UK; 2017. pp. 228-233, DOI: 10.1109/ICORR.2017.8009251
19.Yao S, Zhuang Y, Li Z, Song R. Adaptive admittance control for an ankle exoskeleton using an EMG-driven musculoskeletal model. Frontiers in Neurorobotics. 10 Apr 2018;12:16. DOI: 10.3389/fnbot.2018.00016. PMID: 29692719; PMCID: PMC5902778
20.Ma C, Zhuang Y, Yao S, Song R. Admittance control based on emgdriven musculoskeletal model improves the human-robot synchronization. IEEE Transactions on Industrial Informatics. 2019;152:1211-1218. DOI: 10.1109/TII.2018.2875729
21.Rodrigues-Carvalho C et al. Benchmarking the effects on human–exoskeleton interaction of trajectory, admittance and emg-triggered exoskeleton movement control. Sensors. 2023;23:791. DOI: 10.3390/s23020791
22.Durandau G, Farina D, Asín-Prieto G, Dimbwadyo-Terrer I, Lerma-Lara S, Pons JL, et al. Voluntary control of wearable robotic exoskeletons by patients with paresis via neuromechanical modelling. Journal of Neuroengineering and Rehabilitation. 2019;16:91
23.Menga G, Ghirardi M. Modelling, simulation and control of the walking of biped robotic devices—Part III: Turning while walking. Inventions. 2016;1(1):8
24.Menga G, Ghirardi M. Modelling, simulation and control of the walking of biped robotic devices—Part II: Rectilinear walking. Inventions. 2016;1(1):8
25.Sentis L, Khatib O. Synthesis of whole-body behaviors through hierarchical control of behavioral primitives. International Journal of Humanoid Robotics. 2005;2(4):505-518
26.Choi Y, Kim D, Oh Y, You B-J. Posture/walking control for humanoid robot based on kinematic resolution of CoM jacobian with embedded motion. IEEE Transactions on Robotics. Dec. 2007;23(6):1285-1293. DOI: 10.1109/TRO.2007.904907
27.Alessandro C, Delis I, Nori F, Panzeri S, Berret B. Muscle synergies in neuroscience and robotics: From input–space to task–space perspectives. Frontiers in Computational Neuroscience. 2013;7:43
28.Bahadur R. Ur Rehman S, Khattak S. Synergy estimation for myoelectric control using regularized NMF. In: 2019 41st Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). 2019. pp. 2649-2652. Available from: https://api.semanticscholar.org/CorpusID:204229568
29.Katsavriasa E, Primetisb E, Karandreasa N. Iliopsoas: A new electromyographic technique and normal motor unit action potential values. Clinical Neurophysiology. 2005;116(2005):2528-2532
Notes
An over-parameterized network explains the data it is trained on well, but not so well the validation data or new data when it will be used in the future.
Written By
Giuseppe Menga, Jie Geng and Massimo Mancin
Reviewed: 11 June 2024Published: 11 September 2024
Open access peer-reviewed chapter - ONLINE FIRST
