Method for controlling a movement behaviour of an artificial joint

Abstract

The invention relates to a method for controlling a movement behaviour of an artificial joint which has an upper part and a lower part, mounted thereon to pivot about a pivot axis, between which upper and lower part a device for influencing the pivotability or pivoting of the upper part relative to the lower part is arranged, which device is coupled to a control unit in which a rule set is stored and which activates, deactivates or modulates the device on the basis of input values for the rule set in order to influence the pivoting or pivotability, wherein sensor values of at least one sensor arranged on the upper part or lower part, which are detected during the use of the artificial joint, are supplied to at least one machine-learning-based estimating method which continuously calculates an estimate value for a kinetic or kinematic parameter or an expected kinetic or kinematic parameter from the sensor values and this estimate value is supplied to the rule set as an input value and is used therein as a criterion for activation, deactivation or modulation.

Description

[0001] Method for controlling the movement behavior of an artificial joint

[0002] The invention relates to a method for controlling the movement behavior of an artificial joint, in particular an artificial knee joint, which has an upper part and a lower part pivotably mounted thereon about a pivot axis. Between the upper part and lower part, a device for influencing the pivotability or the pivoting of the upper part relative to the lower part is arranged. The device is coupled to a control device in which a rule set is stored and which activates, deactivates, or modulates the device based on input values ​​for the rule set in order to influence the pivoting or pivotability. The rule set contains methods and control parameters for controlling the artificial joint.

[0003] Artificial joints, particularly artificial knee joints, are found in prostheses and orthoses. Prostheses replace missing limbs in terms of their function and, where appropriate, their external appearance. Orthoses are applied to limbs and serve to guide, limit, and, where appropriate, influence the movement of a natural limb. Orthoses have orthotic joints arranged or formed between an upper and lower part. The upper and lower parts each have fastening devices for attaching the orthosis to the limb. Prostheses have fastening devices with which the prosthesis can be secured to the limb stump or the patient.Between the upper and lower parts of an artificial joint, appropriate devices such as dampers or drives are arranged to influence the pivoting movement or the pivotability. These devices are coupled to a control device via which the dampers or drives are activated, deactivated, or their behavior is modified. Dampers can, for example, be designed as purely passive devices such as linear hydraulics, rotary hydraulics, or magnetorheological dampers. Mechanical brakes can also influence the pivotability or the pivoting movement of the upper part relative to the lower part. Drives are particularly considered to be electric motors and other energy storage devices that can initiate or support a movement or even counteract a movement in order to slow down a pivoting movement.With an appropriate circuit, it is also possible to lock the joint via drives and thus eliminate the ability to pivot.

[0004] The control device activates, deactivates, or modulates the device for influencing the pivoting or pivotability, for example, based on sensor data transmitted to the control device. Sensors are arranged on the artificial joint or on attachments such as the prosthetic socket, distal prosthetic component, or orthotic splint. The sensors can also be arranged on the limb of the treated side or on the contralateral side.

[0005] To control the change in resistance, state machines, for example, are used, which are stored in the control device. A rule set can contain several state machines that are dynamically activated depending on the situation. Based on the sensor data, the state of the prosthesis or orthosis is derived and how an adjustment device, for example for a valve, must be activated or deactivated to generate a specific movement behavior. In a hydraulic resistance device, for example, valves are fully or partially closed to change the cross-section of a fluidic connection in order to influence the corresponding movement of a joint. A control system for a prosthetic knee joint with a state machine is described in EP 549 855 B1.

[0006] DE 10 2020 111 535 A1 discloses a method for controlling at least one actuator of an orthopedic device using an electronic control device. The control device is coupled to the actuator and at least one sensor and has at least one electronic processor for processing sensor data. At least one state machine is stored in the control device, in which states of the orthopedic device and state transitions of the actuator are determined. Furthermore, a classification is stored in the control device, in which sensor data and / or states are automatically classified using a classification method. The classification method and the state machine can be used in combination. Based on the classification and the states, a decision is made as a control signal regarding the manner in which the actuator is activated or deactivated.

[0007] CN 113520683 A discloses a control system for a lower extremity prosthesis and a method for controlling the prosthesis, wherein the prosthesis comprises a control motor for a prosthetic knee joint, a control motor for an ankle joint, a connecting rod, and a sleeve. Gait information of a healthy person is collected under various conditions via an IMU, and a training dataset is created from this. A neural network model is created and trained with the collected data in a simulation environment. The conventional neural network model is implemented in a lower extremity prosthesis in a control device. When the control device receives an input signal from the IMU, an action instruction is output to a joint of the lower extremity, taking into account the trained network model.

[0008] From the article “Knee Angle Estimation based on IMU data and Artificial Neural Networks”, Bennett et al., 29th Southern Biomedical Engineering Conference, 2013, pages 111 and 112, it is known that measuring the knee angle is essential for evaluating gait. The measurement can be performed using an IMU, but due to the indirect measurement, parameters such as knee angle, gait phase, and stance symmetry can only be estimated. The study investigates how an artificial neural network can be used to estimate the knee angle based on acceleration and gyroscope data. It was found that accelerometers are the most effective sensors and that the artificial neural network works best when one IMU is positioned above and one below the knee.Angle sensors, such as knee angle sensors, require installation space on or within the artificial joint and must be mounted, wired, and calibrated separately. The control system requires additional design effort to receive and process the angle data, increasing system complexity.

[0009] The object of the present invention is to provide a method with which, even without angle sensors, it can be detected with sufficient accuracy when and how the pivoting or pivotability must be influenced, whereby the operational effort is as low as possible.

[0010] This object is achieved by a method having the features of the main claim. Advantageous embodiments and further developments of the invention are disclosed in the subclaims, the description, and the figures.

[0011] The method for controlling the movement behavior of an artificial joint, which has an upper part and a lower part pivotally mounted thereon about a pivot axis, between which a device for influencing the pivotability or pivoting of the upper part relative to the lower part is arranged, which device is coupled to a control device in which a rule set is stored and which activates, deactivates or modulates the device on the basis of input values ​​for the rule set in order to influence the pivoting or pivotability, is characterized in that sensor values, e.g. from an IMU arranged on the upper part or the lower part or from IMUs arranged on the upper part and the lower part, which are recorded during use of the artificial joint, are passed to at least one machine learning-based estimation method (MLSV), e.g.an artificial neural network, which calculates an estimated value for a kinetic or kinematic parameter, e.g. joint angle, or an expected kinetic or kinematic parameter from the sensor values ​​and this estimated value is fed to the rule set as an input value, e.g. for the joint angle signal, and is used therein as a criterion for the activation, deactivation, or modulation of the device for influencing the pivotability or pivoting. In one embodiment, it is provided that several identical or different estimation methods based on machine learning (MLSV) determine several estimated values ​​for different kinetic or kinematic parameters and feed them to the rule set. From the data of the sensor or sensors, e.g.The joint angle and, if applicable, another value or values ​​derived from it are continuously calculated using the machine learning-based estimation method(s) of the IMU(s) located on the upper or lower part of the artificial joint or the associated components and limbs. Instead of positioning a direct angle sensor at the joint, such as a knee angle sensor, which requires wiring and calibration, the proposed method creates a virtual joint angle sensor that provides a calculated value or estimated value for the joint angle, in particular the knee angle.The virtual sensor in the machine learning-based estimation method (MLSV) is based on an evaluation of the data from the sensor(s), in particular from one or more IMUs, whereby the machine learning-based estimation method (MLSV) has been trained in particular to calculate the joint angle. The method is not limited to artificial joint devices of the lower extremities, in particular artificial knee joints or ankle joints, but can also be used for other artificial joints, for example, a hip joint or joints of the upper extremities, such as an elbow joint, shoulder joint, or wrist.

[0012] A machine learning-based estimation method (MLSV) can, for example, be implemented as an artificial neural network (ANN), such as a multilayer perceptron, or can include an ANN. Alternatively, black-box models and regression methods are suitable for this purpose. In these models, internal model and calculation parameters are optimized in a training process using previously acquired data. One use case for an MLSV is estimating a knee angle from data from an IMU mounted on the upper or lower part. The training data can be obtained using a system that has both a knee angle sensor and an IMU. Using this data, the MLSV is trained to estimate the knee angle from the IMU data, using the data from the additional knee angle sensor as ground truth.Thus, the "ground truth" contains a reference result for the training process, in which the MLSV learns to estimate the reference result using only the available source data, e.g., IMU data. Once trained, the MLSV can be used as a virtual sensor to estimate the knee angle. This can replace a physical angle sensor. Alternatively, the MLSV can also be trained using unsupervised learning methods.

[0013] Such calculation or determination of a joint angle or knee angle without a direct joint angle sensor is helpful for trip protection in artificial knee joints. Particularly when the flexion resistance and extension resistance can be adjusted separately, trip protection can be implemented in artificial knee joints controlled by a microprocessor-based control unit by increasing the flexion resistance at the time of reversal of movement in the swing phase. This time is also reliably determined using the calculated or estimated value for the knee angle or knee angular velocity calculated in the MLSV. Furthermore, the reversal of movement in the swing phase is a clear and specific situation that is relatively easy to recognize, and therefore inaccuracies in the calculation or estimation can be easily compensated for or are negligible.

[0014] The joint angle input value is thus provided without directly detecting the joint angle via an angle sensor, which results in significantly lower operational effort and less effort for assembly and calibration.

[0015] In one embodiment, the machine learning-based estimation method continuously determines multiple estimated values ​​for one or more kinematic or kinetic parameters, so that, for example, not only the joint angle but also a load, a load profile, accelerations, the spatial orientation of a component with respect to gravity, or the like can be calculated without having to directly measure the desired parameter. In one embodiment, the sensor values ​​are determined by at least one IMU, for example, to estimate the kinetic or kinematic parameter(s) of the joint angle between the upper part and the lower part using the MLSV(s). Multiple identical or different MLSVs can also be used to determine multiple estimated values ​​for different kinematic or kinetic parameters, which then form the basis for the further method.If several estimated values ​​for different kinematic or kinetic parameters are available, these are made available to the rule set as input parameters.

[0016] In one embodiment, additional sensor data and / or state variables are fed into the rule set as input values, so that these additional sensor data and / or state variables are used as criteria for activating, deactivating, or modulating the device for influencing the pivotability or pivoting of the upper part relative to the lower part. This is particularly advantageous in safety-critical cases, since any inaccuracies in the MLSV can be compensated or offset by a plausibility check based on the additional sensor data. Therefore, in one embodiment, safety-critical activations, deactivations, or modulations are not carried out exclusively based on the results of the machine learning-based estimation method, but are secured by additional parameters, measurements, or calculations.

[0017] To improve the accuracy of the calculated value and thus the quality of the input variables for the rule set, the MLSV is supplied with pre-acquired sensor data from a database and trained. The pre-acquired data, stored in the database, makes it easier for the MLSV to estimate the probability of a given situation. The database can be continuously updated and linked to the MLSV to achieve a larger data base and greater accuracy.

[0018] In one embodiment, the database is updated during operation to achieve real-time optimization. Furthermore, inactivity periods, such as charging or non-use, are ideal for refinements and updates to the MLSV, as the system can be updated based on additional data. This data can be collected during operation, thus containing patient-specific information. Alternatively, the data can be made available centrally by the manufacturer, e.g., via the internet as part of a "field update."

[0019] In a further development, the expected joint angle is fed into the rule set with a lead time of between 0.001 seconds and 1 second. This allows the control system to intervene early, but not too early. The MLSV can make predictions about the expected behavior of the upper part relative to the lower part, as well as about the development of the current situation, making it possible to react more quickly to possible actual changes. The prediction is made possible and more accurate because the MLSV can predict the probability of a future movement or future movement behavior with a high degree of accuracy based on the data from the database.Based on these probabilities or estimated values, a corresponding command is then sent or prepared via the control device to adjust the corresponding behavior, i.e., to adjust the pivoting or pivotability of the upper part relative to the lower part about the pivot axis. From the sensor values ​​of the IMU, the probabilities of a movement situation or a state of the joint are calculated in the control device or the MLSV and then fed into the rule set as an input value. The MLSV thus forms a virtual sensor that determines exactly one variable, for example, the knee angle, and then transmits this variable as an input to the control device.In addition to the knee angle, other variables can also be calculated or predicted, such as forces, moments, ground conditions, or gradients of the subsurface. These variables are also calculated by the MLSV with the corresponding probability of occurrence and fed into the control system as an input variable. These input variables are then processed in the respective rule set and serve as the basis for changing or influencing the pivoting resistance of the artificial joint. In one embodiment, sensor values ​​from previous journals or time periods can also be fed into the machine learning-based estimation process to increase the accuracy of determining the estimated value and, in particular, to provide the MLSV with information on the movement history.

[0020] The machine learning-based estimation method(s) is / are based in one embodiment on Gaussian processes, whereby the MLSV(s) provides, in addition to the estimated values ​​for the kinematic or kinetic parameter(s), information on the confidence interval for the respective estimated value, which is then used in the subsequent estimation. The confidence interval for the respective estimated value is used, for example, as a criterion for activation, deactivation, or modulation and, if necessary, transmitted together with the respectively determined estimated value and, if applicable, the rule set. This probabilistic approach uses kernel functions and, in addition to the estimated value, also provides the associated confidence interval, thus providing information on the quality of the estimate.This additional information is of particular interest for the control of orthopedic facilities, as it allows the weighting of the influence of the estimate on the control of the artificial joint to be determined in the rule set. A small confidence interval corresponds to good estimation accuracy and thus a high degree of confidence in the estimate. D, meaning that the estimate can be incorporated into the control with a higher weighting than a less reliable estimate with a large confidence interval.

[0021] Estimates of the information on the ground conditions and gradient of the subsurface are particularly advantageous for controlling artificial ankle joints or prosthetic feet. For example, the foot position at the end of the swing phase can be modulated depending on the estimated gradient of the subsurface. An estimate of the overcome height difference is helpful for optimizing behavior on stairs. Artificial knee joints can modulate the resistance or support behavior depending on these parameters, e.g., when walking downhill, they can offer increased flexion resistance on increasing gradients. When walking uphill, the support moment can be adapted to the difference in height to be overcome. All of this additional information, which is beneficial for optimizing the support performance of artificial joints, can be determined as estimates by appropriately trained MLSVs. Thus, it can be useful to control an artificial joint using several, sometimesto implement MLSVs based on different methods to estimate different parameters (P). For example, a soil quality parameter can be estimated using an ANN, while simultaneously estimating a joint angle using a Gaussian process-based method.

[0022] Exemplary embodiments of the invention are explained in more detail below with reference to the figures. They show:

[0023] Figure 1 - a schematic view of a leg prosthesis;

[0024] Figure 2 - a schematic view of a leg prosthesis in a bent position;

[0025] Figure 3 - a flow chart;

[0026] Figure 4 - a diagram of a knee angle estimation;

[0027] Figure 5 - a schematic view of a leg orthosis;

[0028] Figure 6 - a graph of a parameter regression; and

[0029] Figure 7 - a schematic representation of a prosthetic ankle joint.

[0030] Figure 1 shows a schematic representation of a prosthetic knee joint as part of a leg prosthesis. The prosthetic knee joint has an upper part 10 and a lower part 20, which are pivotally mounted on one another about a pivot axis 15. A prosthetic foot 60 is arranged at the distal end of the lower part 20. In the embodiment as a prosthetic leg according to Figure 1, a prosthetic socket or other device for receiving a femoral stump or for securing it to a person is arranged or formed on the upper part 10.

[0031] Between the upper part 10 and the lower part 20, a device 30 for influencing the pivotability or pivoting of the upper part 10 relative to the lower part 20 is arranged, which device is designed as a linearly acting hydraulic damper. In the illustrated embodiment, the hydraulic damper is designed with a hydraulic chamber or a cylinder that is arranged or formed in a housing or base body 31. A piston 32 is displaceably mounted in the cylinder. The piston 32 is displaceable along the longitudinal extent of the cylinder and is attached to a piston rod 33 that protrudes from the housing or base body 31. The piston 32 divides the cylinder into chambers that are fluidly connected to one another via a hydraulic line.The base body 31 or the housing can be pivotably mounted on the lower part 20 at a fastening point 23 to prevent the piston 32 from tilting during a pivoting movement of the upper part 10 relative to the lower part 20. The end of the piston rod 33 facing away from the piston 32 is attached to the upper part 10, in the illustrated embodiment to an extension arm to increase the distance from the pivot axis 15, at an upper fastening point 21. During flexion, the piston 32 is pushed downward, so that the volume of a flexion chamber decreases. Correspondingly, the volume of an extension chamber increases, reduced by the volume of the retracting piston rod 33.An electric motor can be arranged in the housing 31 to generate pressure within one of the chambers. This motor drives a pump (not shown) to apply pressure to the hydraulic fluid within one of the two chambers and thereby move the piston 32 within the cylinder in one direction or the other. This causes a flexion or extension movement of the orthopedic device in the form of the prosthetic leg. The electric motor for driving the pump is an option that can be used in one embodiment in combination with the linear damper 30. In principle, a drive or motor is not necessary for a passive prosthetic knee joint.An alternative embodiment of the device 30 provides, instead of a linear damper, in particular a linear hydraulic damper, a rotary damper, in particular a rotary hydraulic damper, a magnetorheological resistance device, or an electric motor in generator mode. A combination of several of the aforementioned resistance devices or drives is also implemented in one embodiment. As an alternative to the hydraulic damper, the device 30 can also include a linear actuator, a rotary drive, or a combination of the described technologies.

[0032] Arranged within or on the housing 31 is an actuator 34, which is coupled to at least one control valve 35, via which the hydraulic resistance in the device 30 can be varied. The actuator 34 is coupled to a control device 40, which activates, deactivates, or modulates the actuator 34 based on sensor values ​​in order to provide an adapted resistance and, if necessary, a hydraulic locking mechanism. If the device 30 is configured as a magnetorheological resistance device, the resistances are changed by activating, deactivating, or modulating a magnetic field; the actuator 34 is then the electromagnet or the magnetic coil. If the device 30 also includes active drives, the control device 40 also provides commands for the active energy output by the drive.

[0033] A sensor 50 for detecting the spatial orientation of the lower part 20 and the upper part 10 is arranged on both the upper part 10 and the lower part 20. In particular, the sensor 50 for detecting the spatial orientation is arranged only on the upper part 10. This sensor 50, which is designed as an IMU (inertial measurement unit), determines the solid angle or the absolute angle to a fixed spatial orientation, for example the direction of gravity, during use of the prosthetic knee joint. In addition to detecting spatial positions, the IMU, as sensor 50, can also detect other status data, in particular status data relating to the artificial knee joint. Status data includes, in particular, positions, angular positions, speeds, accelerations, forces, as well as their progressions or changes.The determined solid angle of the upper part 10 and / or the lower part 20 or another state variable is transmitted to the control device 40 as an input variable. The control device 40 modulates, activates, or deactivates the actuator 34 in order to otherwise change the flow resistance in the device 30 in the form of a hydraulic damper, the viscosity, the braking force, or the force counteracting the flexion movement. In order to drive the actuator 34, it is assigned an energy storage device, particularly in the form of an accumulator. The energy storage device can be arranged directly next to the actuator 34 or at another location in the orthopedic device where more space is available or where this appears advantageous due to the weight distribution.In addition to electromechanical drive units with an accumulator or a battery as energy storage, mechanical energy storage devices such as springs or flywheels are also provided in certain embodiments.

[0034] Furthermore, the control device 40 and at least one further sensor 50 can be arranged on the prosthesis foot 60. All sensors arranged on the prosthesis or orthosis are coupled to a control device 40, and their sensor values ​​serve as the basis for controlling the actuator 34 of the device 30 if the latter is designed as a damper, or as input signals for a motor control if the device 30 is designed as a motor. In the case of magnetorheological damping, the sensor values ​​serve to control the magnetic field or its variation.On the basis of the sensor data, in particular the spatial positions as well as position data and data on the load, orientation, acceleration and / or deformation of other components, the actuator 34 is controlled, for example in order to reduce or increase a pivoting resistance, to limit an end stop and / or to generate or support a relative movement between the upper part 10 and the lower part 20.

[0035] Figure 2 shows a schematic representation of a prosthetic knee joint with an upper part 10, a lower part 20, and a device 30 for influencing the pivotability or pivoting of the component 10 relative to the lower part 20 between the upper part 10 and the lower part 20 in a flexed position. The upper part 10 can be pivoted relative to the lower part 20 about the pivot axis 15 against the resistance of the device 30 and has an IMU as a sensor 50. In this process, a joint angle α, in the illustrated embodiment the knee angle α, is changed between the upper part 10 and the lower part 20. In the illustrated embodiment, the joint angle α is measured on the front side between the upper part 10 and the lower part 20. In a fully extended position, the joint angle α is 0°; with increasing flexion, the joint angle α increases and corresponds to the pivot angle.In addition, at least one sensor 50 is arranged on the prosthetic foot 60 on the lower part 20, which sensor is in turn coupled to the control device 40 and which detects a load on the foot part 60 or a pivoting or position of the foot part 60 in space or relative to the lower part 20.

[0036] Figure 3 shows a flowchart of the control system. Sensor values ​​from an IMU as sensor 50, for example, acceleration values ​​and / or spatial orientations of IMUs arranged on the upper part 10 and / or the lower part 20 of an orthopedic joint device, for example, a prosthetic knee joint or an orthotic knee joint, are transmitted to a machine learning-based estimation method (MLSV). The MLSV can be trained both for regression tasks for estimating output variables or probabilities and for classification tasks. The sensor values ​​from the IMU 50 are evaluated within the MLSV with regard to a possible knee angle α. The evaluation within the machine learning-based estimation method (MLSV) results in a calculated value or estimated value K, which is processed in the control device 40 instead of a direct knee angle sensor signal.The estimated value K serves as an input value or an input variable for the rule set stored in the control device 40. In addition to the estimated value K, further sensor values ​​from sensors 55 that have not been evaluated by the MLSV can be fed to the control device 40 with the rule set. Based on the input values, the control device 40 itself determines a control command for the actuator to increase or decrease a resistance or to activate or deactivate a drive. Within the MLSV, a probability can also be calculated regarding the current situation of the user of the prosthesis or orthosis, or a classification value corresponding to a specific situation.The estimated value K for the knee angle α is therefore not measured directly via a knee angle sensor, but is preferably determined by the artificial intelligence of the MLSV based on one or more sensor values ​​from the IMU arranged on the upper part 10 or the lower part 20. For this purpose, data from a database for training the artificial intelligence or the MLSV is used in the artificial intelligence, in an artificial neural network (KNN), or another MLSV. In addition to the estimated value K for the knee angle α, other variables or input values ​​for the control device 40 can also be calculated or estimated based on the sensor values ​​from the IMU and, if applicable, other sensors transmitted to the MLSV, for example forces or moments that may occur or be expected due to movement sequences.Based on the evaluation of the sensor data within the MLSV, it is possible to make predictions or probabilities about the future course of variables or parameters, since the MLSV performs the evaluation from previous data and data histories. Using data histories and calculating probabilities makes it possible to observe and anticipate the further course of the data curve of the variable of the artificial joint with a corresponding probability. However, the longer the time span to be predicted or anticipated, the lower the probability of the prediction occurring. Therefore, the lead time with which the MLSV supplies the control device 40 with the expected joint angle value a or another expected variable is limited.In particular, the limit is between 0.001 seconds and 1 second, in order to, on the one hand, still have a noticeable effect of the pre-setting and, on the other hand, to avoid excessive changes to the setting of the device 30 that are not justified by the actual situation in or on the joint. By using the MLSV, the use of an estimated value K or calculated value determined based on measured historical data from other orthopedic devices, future movement sequences and loads can be estimated and used to improve control. The MLSV can also be supplied with additional sensor signals other than IMU signals in order to improve the prediction or the calculation accuracy for determining the variable to be determined.

[0037] With such a method, the control functions without a direct joint angle sensor and the rule set within the control device 40 is supplied exclusively with data from one or more IMUs.

[0038] Figure 4 shows the result of an estimation or calculation of the knee angle α using the MLSV, in this case an artificial neural network, based on the curve αP predicted by the MLSV as a dotted line compared to a directly measured knee angle αR as a solid line. The two curves show essentially the same course. For the calculated knee angle αP, which was calculated only on the basis of raw data from the IMU, i.e., acceleration and gyroscope values, there are deviations between the maximum values ​​at small knee angles α and the actual values ​​αR measured directly with a knee angle sensor. Over the course of the steps, the two curves αP and αR become increasingly similar. In addition, the result can be further improved by incorporating additional sensor signals, for example, spatial orientation.Especially when detecting significant events such as the reversal of movement of a lower leg at the end of a swing phase in an orthotic or prosthetic knee joint, the accuracy of the MLSV measurement is sufficient. Indirect detection without direct knee angle sensors is advantageous in this case, as an additional sensor is unnecessary.

[0039] Figure 5 shows a schematic representation of an orthosis with a basic structure that corresponds to the prosthesis according to Figure 1. The connection between the artificial joint and the leg is made in this case via connecting devices 101, 201. One difference between the embodiment according to Figure 1 and the embodiment according to Figure 5 is that, according to Figure 5, a drive 70 is additionally provided, in which an electric motor 70 is coupled to a pulley, optionally via a gear mechanism. Depending on the direction of rotation of the motor 70, flexion or extension of the knee joint can then be effected or supported via a V-belt or toothed belt. The embodiment with the drive with an electric motor 70 via a mechanical power transmission device and parallel damping via a hydraulic damper 30 can also be used with a prosthetic knee joint.As already explained in Figure 1, the resistance device in an orthosis can also be formed by a motor, e.g. in generator mode.

[0040] The direct mechanical coupling of the electric motor as a resistance device 30 to the upper part 10 and the lower part 20 can be achieved via a power transmission device, for example, via a spindle drive, so that instead of a piston rod 33, a spindle is retracted or extended from the housing 31 by turning a spindle nut driven by the electric motor. In another embodiment, the motor as a resistance device is coupled to the upper part 10 and the lower part 20 via a gear device, for example, via a planetary gear, in order to effect or decelerate a displacement of the upper part 10 relative to the lower part 20.

[0041] Furthermore, the control device 40 and at least one IMU as a sensor 50 are arranged on the prosthesis or orthosis. The angle between the upper part 10 and the lower part 20 is determined by evaluating the sensor data from two spatial position sensors or IMU 50. All sensors arranged on the prosthesis or orthosis are coupled to a control device 40, and their sensor values ​​serve as the basis for controlling the actuator 34 of the resistance devices 30 if this is designed as a damper, or as input signals for a motor control for the electric motor 70 if the resistance device 30 is designed as a motor. In the case of magnetorheological damping, the sensor values ​​serve to control the magnetic field or its variation.Based on the sensor data, in particular the spatial positions and / or angular positions as well as position data and data on the load, orientation, acceleration and / or deformation of other components, the actuator 34 is controlled, or the electric motor 70 is activated, deactivated or modulated, for example in order to reduce or increase a pivoting resistance, to limit an end stop and / or to generate or support a relative movement between the upper part 10 and the lower part 20.

[0042] Figure 6 shows, as an example, the result of a regression of a parameter plotted on the Y-axis from an input value plotted on the X-axis in the form of a solid line. The areas ±0 and ±2o plotted around the solid line indicate the respective confidence intervals. The input data, which is supplied in the form of training data, is plotted as points on the solid line. In the area of ​​the training data, the quality of the estimate is better, the confidence intervals are smaller, which means that with an increased amount of training data, the quality of the calculated data improves. Figure 7 schematically shows a prosthetic ankle joint in which the upper part 10 is a lower leg socket and the lower part 20 is the prosthetic foot.Sensors 50 are arranged on both the lower part 20 and the upper part 10. These sensors are coupled to the control device 40, via which the actuator 30 or the resistance device is influenced accordingly. For example, the resistance against dorsiflexion and / or plantar flexion is adjusted depending on an estimated value for a kinetic or kinematic parameter. The estimated value is fed into the rule set and can, for example, also relate to the ground inclination.

Claims

Patent claims 1. A method for controlling the movement behavior of an artificial joint, which has an upper part (10) and a lower part (20) pivotably mounted thereon about a pivot axis (15), between which a device (30) for influencing the pivotability or the pivoting of the upper part (10) relative to the lower part (20) is arranged, which is coupled to a control device (40) in which a rule set is stored and which activates, deactivates, or modulates the device (30) on the basis of input values ​​for the rule set in order to influence the pivoting or pivotability, characterized in that sensor values ​​of at least one sensor (50) arranged on the upper part (10) or the lower part (20), which are recorded during use of the artificial joint, are fed to at least one machine learning-based estimation method (MLSV).which continuously calculates an estimated value (K) for a kinetic or kinematic parameter (a) or an expected kinetic or kinematic parameter (a) from the sensor values ​​and this estimated value (K) is fed as an input value to the rule set and is used therein as a criterion for activation, deactivation or modulation.

2. Method according to claim 1, characterized in that the machine learning-based estimation method (MLSV) continuously determines several estimated values ​​(K) for several kinematic or kinetic parameters (a).

3. Method according to claim 1 or 2, characterized in that the sensor values ​​are determined by at least one IMU.

4. Method according to one of the preceding claims, characterized in that the kinetic or kinematic parameter (a) includes or represents a joint angle between the upper part (10) and the lower part (20).

5. Method according to claim 4, characterized in that the joint angle input value is provided via an angle sensor without direct detection of the joint angle (a).

6. Method according to one of the preceding claims, characterized in that additional sensor data and / or state variables are supplied to the rule set as input values ​​and are used as criteria for activation, deactivation or modulation.

7. The method according to claim 6, characterized in that safety-critical activations, deactivations or modulations are not carried out exclusively based on the results of the machine learning-based estimation method (MLSV).

8. Method according to one of the preceding claims, characterized in that the machine learning-based estimation method (MLSV) is supplied with previously obtained sensor data from a database and trained.

9. The method according to claim 8, characterized in that the machine learning-based estimation method (MLSV) is carried out with a defined data set.

10. Method according to one of the preceding claims, characterized in that the machine learning-based estimation method (MLSV) is updated during operation with current data obtained during operation.

11. Method according to one of the preceding claims, characterized in that the machine learning-based estimation method (MLSV) includes an artificial neural network (ANN).

12. Method according to one of the preceding claims, characterized in that the machine learning-based estimation method (MLSV) is based on a regression method or a parameterized black box model.

13. Method according to one of the preceding claims, characterized in that the expected joint angle (a) is fed to the rule set with a lead time between 0.001 s and 1 s.

14. Method according to one of the preceding claims, characterized in that probabilities of a movement situation or a state of the joint are calculated from the sensor values ​​in the machine learning-based estimation method (MLSV) and fed to the rule set as input value.

15. Method according to one of the preceding claims, characterized in that sensor values ​​from previous time steps are also fed to the machine learning-based estimation method (MLSV).

16. Method according to one of the preceding claims, characterized in that the machine learning-based estimation method (M LSV) is based on Gaussian processes and provides information on a confidence interval (Kia) for the respective estimated value (a).

17. The method according to claim 16, characterized in that in addition to the determined estimated value (a), the associated confidence interval (Kia) is also supplied to the rule set and used as a criterion for activation, deactivation or modulation.

18. The method according to claim 17, characterized in that several estimated values ​​(a) are transmitted to the rule set together with associated confidence intervals (Kia).

19. Method according to one of the preceding claims, characterized in that several estimated values ​​(a) for different kinematic or kinetic Parameters (P) are made available to the rule set as input parameters.

Images