Method for autonomous calibration of electromechanical prosthesis control device
Patent Information
- Authority / Receiving Office
- RU · RU
- Patent Type
- Patents
- Current Assignee / Owner
- OBSHCHESTVO S OGRANICHENNOJ OTVETSTVENNOSTYU MOTORIKA
- Filing Date
- 2025-04-25
- Publication Date
- 2026-07-08
AI Technical Summary
Existing electromechanical prosthesis calibration methods are inefficient due to high resource intensity and computational demands, leading to reduced accuracy and prolonged calibration times, and increased power consumption, which negatively impacts user experience.
An autonomous calibration method using electromyographic sensors, real-time data collection, and batch gradient descent to train a machine learning model, reducing resource intensity and optimizing computational algorithms for efficient gesture recognition.
The method improves the accuracy and efficiency of prosthetic control by minimizing resource consumption, reducing processing delays, and enhancing gesture recognition, resulting in a more reliable and user-friendly prosthesis.
Smart Images
Figure 00000001 
Figure 00000002
Abstract
Description
[0001] The invention relates to methods for adapting a gesture recognition system integrated with an electromechanical prosthetic limb, manipulator or virtual device to the characteristics of myographic signals received from a user, and can be used in the medical industry.
[0002] Currently, electromechanical prosthetic technologies for missing limbs are rapidly developing. Typically, an electromechanical prosthesis is controlled by its controller based on signals characterizing muscle condition obtained via myographic sensors. Physiological factors such as the user's muscle tone and strength, sweating, skin condition, water and fat content, and the fit of the prosthetic socket on the user's residual limb can change over time. Changes in these parameters can significantly alter myographic sensor readings and, consequently, reduce the accuracy of electromechanical prosthesis control. Environmental factors such as temperature and humidity can also influence sensor readings.In this regard, periodic calibration of the electromechanical prosthesis control system may be required to adapt the control system to the user's condition.
[0003] The process of calibrating the control device of an electromechanical prosthesis involves processing myographic data, which is quite large and contains overlapping information. Machine learning models are currently successfully used to process large data sets.
[0004] One example of such use of a machine learning model is a prototype method for calibrating a control device for a user's electromechanical prosthesis. This involves attaching electromyographic sensors to the user's residual limb and installing the prosthesis on it. The prosthesis then, under the control of a controller, performs a sequence of gestures for visual observation by the user, after which the user is prompted to perform these gestures. As the user performs gestures, the electromyographic sensors collect data on the electrical activity of the muscles, after which the prosthesis controller processes the entire volume of registered myographic data using a machine learning model, as a result of which the model is trained to control the prosthesis [US 2014032462 A1, publication date: 30.01.2014].
[0005] A drawback of the prototype is the low efficiency of the calibration method for the user's electromechanical prosthesis control device. This is due to the lack of optimization of controller resources when processing large volumes of myographic data and the lack of optimization of the computational algorithms. Specifically, this lack of optimization results in the model training using the entire volume of myographic data recorded during the user's gesture sequence. Processing this entire dataset requires significant computing resources and large amounts of RAM. This necessitates the use of computing devices with a powerful microcontroller or processor and large memory capacity.The use of such computing devices increases the power consumption of the computing device underlying the prosthesis control system, which may result in a reduction in the battery life of the prosthesis, requiring more frequent recharging, which negatively impacts the user experience, or, while maintaining the battery life, requiring the use of batteries with a higher capacity, which will lead to an increase in the weight of the prosthesis.
[0006] However, even with powerful microcontrollers, there is a risk of errors due to RAM overload when attempting to process a large array of myographic data. This can lead to abnormal situations during calibration and, consequently, to reduced gesture recognition accuracy by the prosthetic controller in the future. Furthermore, processing a large array of myographic data, even with powerful microcontrollers, increases the time it takes to calibrate the prosthetic control device.
[0007] Therefore, it is necessary to develop a method for calibrating the control device of the user's electromechanical prosthesis, eliminating the existing shortcomings.
[0008] The technical problem that the invention is aimed at solving is the need to eliminate the shortcomings of the existing method for calibrating the control device of the user's electromechanical prosthesis.
[0009] The technical result that the invention is aimed at achieving consists in increasing the efficiency of the method for autonomously calibrating the control device for the user's electromechanical prosthesis by reducing the resource intensity of the algorithms for automatically calibrating the control device for the electromechanical prosthesis.
[0010] The essence of the invention is as follows.
[0011] A method for autonomous calibration of a control device for a user's electromechanical prosthesis, during which the microcontroller executes an automatic calibration protocol, during which:
[0012] - the target variable, which includes the gesture required to be played by the user, is visualized through the output device in accordance with the commands from the microcontroller;
[0013] - the user performs gestures according to the visualization, and the electromyographic sensors attached to the residual part of the user's limb collect muscle electrical activity data in real time at a specified sampling rate;
[0014] - Simultaneously with data collection, the machine learning model contained in the microcontroller is trained, in which the weighting coefficients are initialized to zeros or random numbers, depending on the machine learning model used, for which:
[0015] - the data obtained from the electromyographic sensors are entered into a data batch containing classes of observations with assigned values of signals from the electromyographic sensors, replacing a randomly selected observation of the same class that was already in the batch;
[0016] - The machine learning model is trained in streaming mode using batch gradient descent based on the data in the batch;
[0017] - model weights are updated after each iteration of batch gradient descent;
[0018] - The trained machine learning model is used by the microcontroller to recognize the gestures of the user of the electromechanical prosthesis.
[0019] To implement the method for autonomous calibration of the control device of the user's electromechanical prosthesis, a system may be used that contains electromyographic (EMG) sensors, an output device and a microcontroller, the software of which includes a data pre-processing module, a machine learning model and a command generation module for the output device.
[0020] EMG sensors provide real-time, frequency-controlled readings of muscle electrical activity. EMG sensors can be attached to a cuff or prosthetic socket, ensuring reliable and stable data collection.
[0021] The output device provides visualization of the target variable according to the commands from the microcontroller, and can be represented by a prosthesis installed on the user's stump, or a prosthesis installed separately from the user's stump, or a virtual prosthesis displayed on the display.
[0022] The microcontroller controls the prosthesis and manages the calibration process using software that includes a data preprocessing module, a machine learning model, and an output device command generation module. The data preprocessing module prepares data recorded by the EMG sensors for use in the machine learning model. Specifically, it can perform data augmentation, such as neutral position augmentation, which improves gesture absence recognition and reduces the number of false positives in gesture presence recognition. It can also apply data filtering, such as a high-pass filter, and perform feature extraction, generating and updating data batches used in model training.Preprocessed EMG data and the labeled target variable are fed to the machine learning model as training data. The machine learning model can be initialized with zeros or random numbers. During model training, the model weights are updated after each iteration of batch gradient descent. The output device command generation module generates commands for real or virtual actuators of the output device represented by the prosthesis and also facilitates data exchange with the preprocessing module for labeling the target variable.
[0023] The autonomous calibration process of the user's electromechanical prosthesis control device begins with the installation of electromyography (EMG) sensors on the residual part of the user's limb.
[0024] The microcontroller initiates automatic calibration mode, during which the output device first displays a target gesture, such as a "Pull" gesture, a "Point" gesture, or other gestures. The target gesture is displayed under the control of the output device command generation module. At the same time, the output device command generation module sends data about the displayed gesture to the preprocessing module, thus mapping the target variable.
[0025] The user's gestures are realized by the corresponding muscle contraction of the residual limb, accompanied by the generation of corresponding electrical potentials. The electrical activity of the muscles is recorded by EMG sensors and sent to the pre-processing module. The EMG sensor sampling rate can be as low as 33 Hz. This sampling rate allows for the acquisition of new data and its use for model training in real time, as the data processing time for a single training iteration using batch gradient descent is no more than 30 ms.
[0026] Based on the target variable data received by the pre-processing module, as well as the muscle electrical activity data, this module generates a data batch, which is a data set containing observation classes with assigned signal values from EMG sensors.
[0027] Based on the data batch generated by the pre-processing module, the machine learning model is trained to recognize user gestures.
[0028] As the EMG sensors are sampled at a specified rate, electrical muscle activity data is fed to the preprocessing module, where it is assigned a class based on the target variable data transmitted from the command generation module. This creates a new observation. A new observation of a given class is added to the batch with a certain probability and replaces a random observation of the same class already in the batch. This results in an updated data batch, which is used to further train the model to recognize user gestures. This new training iteration is used, and the entire model training process is based on the principle of batch gradient descent. After each training iteration, the model's weighting coefficients are updated. Each input data batch is used only once, and the batch preceding the updated one is not stored in the microcontroller's memory.This training algorithm allows for a reduction in the resource intensity of the automatic calibration process, thereby increasing the efficiency of the method for autonomously calibrating the control device for the user's electromechanical prosthesis when using microcontrollers with low performance and under conditions of limited memory resources, which are available in control devices for electromechanical prostheses.
[0029] After each training iteration, the model updates its weighting coefficients, allowing it to more accurately recognize user gestures. This, in turn, improves the accuracy of prosthetic control, reducing false positives and incorrectly identified classes. Upon completion of the automatic calibration process, the microcontroller switches to operation mode, where the trained model is used to recognize gestures.
[0030] Additionally, to improve the efficiency of the autonomous calibration method for the control device of an electromechanical prosthesis, the neutral position in which the user does not perform any gesture can be augmented at the data pre-processing stage, thereby improving the quality of recognition of the absence of a gesture and reducing the number of false positive errors in recognition of the presence of a gesture.
[0031] The invention can be made from known materials using known means, which indicates its compliance with the patentability criterion of “industrial applicability”.
[0032] The invention is characterized by a previously unknown set of essential features. The method for autonomously calibrating a user's electromechanical prosthesis control device offers a number of innovative solutions that significantly improve the system's functionality, ensuring its efficiency and reducing resource consumption. Let's examine in detail how each of these features contributes to achieving these goals and eliminates known shortcomings:
[0033] - Electromyographic sensors attached to the user's residual limb collect muscle activity data in real time at a preset sampling rate. This allows the system to quickly process incoming information. This operating scheme reduces the load on the microcontroller and memory, as the system does not require storing large amounts of data for subsequent processing. This also minimizes delays in command recognition and execution, which is critical for the speed and precision of prosthetic control.
[0034] - Using batch gradient descent to train the model allows for data processing in small groups or batches. This approach significantly reduces memory requirements and the load on the microcontroller, as data is processed gradually rather than all at once. This improves system resource efficiency and accelerates the training process, delivering more stable and accurate results. This approach can also reduce computing power consumption, which is especially important for standalone and mobile devices.
[0035] - Regularly updating the model weights after each training iteration allows the system to adapt to changes in the data. This ensures optimal model parameters, which subsequently improves the accuracy of user movement recognition during device operation.
[0036] The set of essential features of the invention significantly improves the efficiency of autonomous calibration of an electromechanical prosthesis control device under conditions of limited microcontroller performance and device memory capacity by reducing the resource intensity of the autonomous calibration method for the electromechanical prosthesis control device. Optimized computational resource utilization, improved initial model setup, and efficient data management eliminate many drawbacks, such as high system load, power consumption, and operational delays. As a result, the system becomes more reliable, accurate, and error-tolerant, significantly improving the user experience and making the prosthesis more functional and user-friendly.
[0037] This ensures the achievement of a technical result consisting in increasing the efficiency of the method for autonomous calibration of the control device of the user's electromechanical prosthesis by reducing the resource intensity of the algorithms for automatic calibration of the control device of the electromechanical prosthesis.
[0038] The invention has a set of essential features previously unknown in the prior art, which indicates its compliance with the patentability criterion of “novelty”.
[0039] The prior art does not disclose any essential distinguishing features of the claimed method for autonomous calibration of a control device for a user’s electromechanical prosthesis, which is why the invention meets the patentability criterion of “inventive step”.
[0040] The invention is explained by the following figures.
[0041] Fig. 1 - System for performing a method for autonomous calibration of a control device for a user's electromechanical prosthesis.
[0042] Fig. 2 - Algorithm for executing the automatic calibration protocol within the framework of the method of autonomous calibration of the control device of the user's electromechanical prosthesis.
[0043] To illustrate the possibility of implementation and a more complete understanding of the essence of the invention, an embodiment of it is presented below, which can be changed or supplemented in any way, while the present invention is in no way limited to the presented embodiment.
[0044] A method for autonomously calibrating a control device for a user's electromechanical prosthesis is implemented by a system that includes electromyographic (EMG) sensors 100, an output device 110 that can be represented by a prosthesis installed on the user's stump, or a prosthesis installed separately from the user's stump, or a virtual hand or prosthesis displayed on a display, and a prosthesis control device represented by a microcontroller with a RAM capacity of 128 KB and including a data preprocessing module 120, a machine learning model 130, and a module 140 for generating commands for the output device.In order to implement the method, a machine learning model 130 is selected in which the weight coefficients are initialized with zeros or random numbers, such as, for example, logistic regression or its variants, a decision tree or their ensembles of the random forest type, or neural networks based on fully connected, recurrent and convolutional layers, etc. The connection of the above-described system components is ensured as follows: data from the EMG sensors 100 and information on the value of the target variable from the module 140 for generating commands for the output device are fed to the input of the pre-processing module 120, data from the pre-processing module 120 are fed to the input of the machine learning model 130, the output of the machine learning model 130 is connected to the input of the module 140 for generating commands for the output device, and commands from the module 140 for generating commands for the output device are fed to the input of the output device 110.
[0045] Before performing the method, a prosthetic socket or cuff with EMG sensors 100 attached to it is installed on the user's stump and they and the output device 110 of the selected type are connected to the microcontroller.
[0046] The method for autonomous calibration of the control device of the user's electromechanical prosthesis is implemented as follows.
[0047] The microcontroller initiates the automatic calibration protocol. When executing the automatic calibration protocol, the output device command generation module 140 sends a command to the output device 110 of the selected type to perform a gesture required to be performed by the user, and simultaneously sends information about the type of gesture required to be performed by the user (e.g., the "Pointing" gesture) to the data preprocessing module 120, which serves as a target variable label.
[0048] After this, the output device 110 performs the required gesture, and then the user performs a gesture associated with the prosthetic gesture. This gesture can either replicate the prosthetic gesture or be any gesture the user can repeatedly perform. The resulting muscle contraction in the remaining limb is accompanied by a change in the electrical potential of the muscles.
[0049] When the user performs gestures, the EMG sensors 100 record the electrical activity data of the muscles and send it to the data pre-processing module 120, and the sampling frequency of the EMG sensors 100 is 33 Hz (the period is 30 ms).
[0050] Module 120 preprocesses the muscle electrical activity data obtained by EMG sensors 100 and, based on this data and the labeled target variable, generates a data batch and sends it to machine learning model 130. The data batch generated by module 120 contains observation classes with assigned signal values from EMG sensors 100.
[0051] Next, based on the batch of data generated by module 120, machine learning model 130 is trained to recognize user gestures.
[0052] As EMG sensors 100 are sampled every 30 ms, the muscle electrical activity data is fed to pre-processing module 120, generating new observations with assigned signal values from EMG sensors 100. Module 120 can perform data augmentation, such as neutral position augmentation, which improves gesture absence recognition and reduces false positives in gesture presence recognition. Module 120 can also apply filters, such as a high-pass filter, and perform feature extraction. A new observation of a certain class is included in the batch with a certain probability and replaces a random observation of the same class already in the batch. The probability of inclusion in the batch is determined by generating a random number, and if this number is higher than a preset threshold, this value is included in the batch.This ensures data diversity in the batch, allowing the model to train on data with a value distribution close to its expected distribution under operating conditions. This results in an updated data batch, which is used to further train model 130 to recognize user gestures. This process involves a new training iteration, and the training process for model 130 is based on the principle of batch gradient descent. After each training iteration, the weights of model 130 are updated. Each data batch input to model 130 is used only once, and the batch preceding the updated one is not stored in the microcontroller's memory.
[0053] The number of data collection cycles in the automatic calibration protocol is set in the microcontroller software. The calibration protocol is executed until the specified number of cycles is completed.
[0054] After completing the automatic calibration protocol, the trained model is used to recognize the gestures of the prosthesis user, and the microcontroller enters the user operation mode.
[0055] Thus, the technical result is achieved, which consists in increasing the efficiency of the method for autonomous calibration of the control device of the user's electromechanical prosthesis by reducing the resource intensity of the algorithms for automatic calibration of the control device of the electromechanical prosthesis.
Claims
1. A method for autonomous calibration of a control device for a user's electromechanical prosthesis, during which the microcontroller executes an automatic calibration protocol, during which: - the target variable, which includes the gesture required to be reproduced by the user, is visualized by means of an output device in accordance with commands from the microcontroller; - the user performs gestures in accordance with the visualization, and electromyographic sensors attached to the residual part of the user's limb collect data on the electrical activity of the muscles in real time at a specified sampling frequency; - simultaneously with data collection, the machine learning model contained in the microcontroller is trained, in which the weighting coefficients are initialized to zeros or random numbers, depending on the machine learning model used, for which: - the data obtained from the electromyographic sensors are entered into a data batch containing classes of observations with assigned values of signals from the electromyographic sensors, replacing a randomly selected observation of the same class that was already in the batch; - the machine learning model is trained in streaming mode using batch gradient descent based on the data in the batch; - the model weights are updated after each iteration of batch gradient descent; - the trained machine learning model is used by the microcontroller to recognize the gestures of the user of the electromechanical prosthesis.
2. The method according to paragraph 1, characterized in that when the microcontroller executes the automatic calibration protocol, the neutral position is augmented.