Wearable human state assessment device

Abstract

The application discloses a wearable human state evaluation device, comprising: an electromyographic signal acquisition module, which is used for collecting electromyographic signals of a user, and comprises five channels; an MCU module, which is connected to the electromyographic signal acquisition module, is used for receiving the electromyographic signals collected by the electromyographic signal acquisition module, and processes the received electromyographic signals; an IMU module, which is used for collecting acceleration data of the user; and a wireless transmission module, which is connected to the MCU module and the IMU module, receives the processed data of the MCU module and the collected acceleration data of the IMU module, and wirelessly sends the data to an upper computer. The wearable human state evaluation device of the application processes electromyographic signals by using non-negative matrix decomposition, thereby reducing data transmission amount, improving data transmission efficiency, reducing system power consumption, and being beneficial to long-time continuous work of the wearable device.

Description

Technical Field

[0001] This invention specifically relates to a wearable human condition assessment device. Background Technology

[0002] Human condition assessment, particularly for fall and balance monitoring, has received widespread attention in recent years. Technological advancements in this field have primarily benefited from the rapid progress of wearable devices and the integration of sensor technology, data analysis methods, and artificial intelligence. However, existing wearable devices suffer from issues such as large data transmission volumes and inaccurate collision detection. Summary of the Invention

[0003] This invention provides a wearable human body condition assessment device that solves at least one of the aforementioned technical problems, specifically employing the following technical solution:

[0004] A wearable human condition assessment device, comprising:

[0005] An electromyography (EMG) signal acquisition module is used to acquire the user's EMG signals, and the EMG signal acquisition module contains five channels;

[0006] The MCU module is connected to the electromyography signal acquisition module and is used to receive the electromyography signals acquired by the electromyography signal acquisition module and process the received electromyography signals.

[0007] The IMU module is used to collect the user's acceleration data;

[0008] A wireless transmission module is connected to the MCU module and the IMU module. The wireless transmission module receives the data processed by the MCU module and the acceleration data collected by the IMU module, and wirelessly transmits them to the host computer.

[0009] Furthermore, the wireless transmission module includes at least one of a WiFi module or a Bluetooth module.

[0010] Furthermore, the wireless transmission module is wirelessly connected to the IMU module.

[0011] Furthermore, the MCU module includes a computing unit and five ADC units, each of which performs analog-to-digital conversion on the five channels of electromyography signals. The computing unit is used to perform NNMF processing on the converted multi-channel electromyography signals.

[0012] Furthermore, the host computer obtains COP and GRF based on the received NNMF-processed data and acceleration data, and determines the user's state based on the obtained COP and GRF.

[0013] Furthermore, the host computer obtains COP and GRF through a pre-trained prediction model.

[0014] Furthermore, the prediction model comprises: an input layer, an LSTM layer, a COH collaborative consistency layer, an acceleration feature layer, a feature fusion layer, and a fully connected output layer;

[0015] Normalized multidimensional surface electromyography time series data are input into the input layer;

[0016] Multidimensional surface electromyography time-series data are input into the LSTM layer to obtain LSTM features;

[0017] Multidimensional surface electromyography time-series data are input into the COH coherence layer to obtain COH features;

[0018] The acceleration feature layer is used to extract acceleration features from the acceleration data;

[0019] The feature fusion layer fuses LSTM features, COH features, and acceleration features to obtain fused features;

[0020] The fused features are input into the fully connected output layer to obtain the predicted COP and GRF.

[0021] Furthermore, the LSTM layer is a BiLSTM layer;

[0022] Multidimensional surface electromyography time-series data are input into the BiLSTM layer to obtain BiLSTM features;

[0023] The feature fusion layer fuses BiLSTM features, COH features, and acceleration features to obtain fused features.

[0024] Furthermore, the BILSTM layer employs either the BILSTM model or the CNN-BILSTM model.

[0025] Furthermore, the Adam optimizer is employed when training the prediction model.

[0026] The advantage of this invention is that the wearable human condition assessment device provided reduces the amount of data transmission, improves data transmission efficiency, and reduces system power consumption by processing electromyographic signals using nonnegative matrix decomposition, which is beneficial for the long-term continuous operation of the wearable device. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 A schematic diagram of a wearable human condition assessment device provided by the present invention;

[0029] Figure 2 This is a schematic diagram of the prediction model in an embodiment of the present invention. Detailed Implementation

[0030] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0031] like Figure 1 The image shows a wearable human condition assessment device (cooperative detection minimum unit) according to this application, which includes: an electromyography signal acquisition module, an MCU module, an IMU module, and a wireless transmission module.

[0032] Specifically, the electromyography (EMG) signal acquisition module is used to acquire the user's EMG signals. The MCU module connects to the EMG signal acquisition module to receive and process the acquired EMG signals. The IMU module acquires the user's acceleration data. The wireless transmission module connects to both the MCU and IMU modules, receiving the processed data from the MCU module and the acceleration data acquired by the IMU module, and wirelessly transmitting it to the host computer. The host computer then determines the user's status based on the received data.

[0033] Conventional collaborative analysis typically relies on multiple sensors (usually 12 or more channels of electromyography (EMG) sensors) to upload raw data to a host computer, either separately or in a unified manner. Separate approaches generally include several independent EMG acquisition and data transmission modules; for example, a 12-channel system typically contains 12 independent modules. This approach requires mounting a large number of sensor modules in different locations on the body, resulting in high data transmission volume, complexity, and inefficiency due to the simultaneous monitoring of muscle data irrelevant to the specific task. Unified approaches generally include a desktop host that acquires surface electrical signals via wired connections and transmits them to the host for processing. While this method avoids mounting multiple sensor modules on the body, it requires long cables that restrict the user's movement. For backpack-style host systems with multiple channels, real-time data transmission and analysis are not feasible.

[0034] Based on the above description, in the embodiments of this application, the electromyography (EMG) signal acquisition module includes five channels, each containing two or three muscle detection electrodes. The MCU module includes a computing unit and five ADC units, which respectively perform analog-to-digital conversion on the EMG signals of the five channels. The EMG signal detected by each channel is sampled by the ADC unit of the main control MCU module to complete the conversion between analog and digital signals. In a specific case, for example, channel 1 acquires the EMG signal of the tibialis anterior (TA) muscle. According to the SENIAM guidelines, in order to reduce the impedance of the skin surface, the skin surface of the muscle to be tested is first cleaned with alcohol; then, the acquisition electrodes are attached to the skin surface of the muscle belly along the direction of the muscle fibers, with the electrodes spaced 2 cm apart, and the reference electrode is attached to the bone or unrelated muscle skin surface. The TA EMG signal acquired by channel 1 is converted from analog to digital by ADC1.

[0035] The computational unit performs Non-Negative Matrix Factorization (NNMF) processing on the converted multi-channel electromyography (EMG) signal. All elements in the basis matrix and coefficient matrix obtained by NNMF decomposition are non-negative, providing a reliable interpretation of the non-negativity of the EMG signal. Therefore, NNMF can decompose the EMG signal into non-negative temporal and spatial structures. By iteratively constraining to minimize the error between the reconstructed matrix and the original matrix, the coordination matrix and activation matrix can be obtained as accurately as possible. The processed signal is transmitted to the wireless transmission module via UART-UART1, and further transmitted to the host computer. It is understood that the wireless transmission module includes at least one of a WiFi module or a Bluetooth module. In the embodiments of this application, the wireless transmission module is wirelessly connected to the IMU module. Preferably, the wearable human condition assessment device of this application also includes a pressure sensing unit, which detects pressure signals and transmits the detected pressure signals to the wireless transmission module.

[0036] NNMF (Neural Non-negativity Matrix) is widely used to extract the spatiotemporal structure of muscle coordination due to its non-negativity constraint property. By iteratively constraining and minimizing the error between the reconstructed signal matrix and the original signal matrix, it can decompose multi-channel signals into weight matrices and co-activation matrices as accurately as possible. The NNMF of multi-channel EEG signals can be represented as a linear combination of co-weights and co-activations.

[0037]

[0038] in It is the synergistic weight matrix, which defines the spatial pattern of muscle synergy. Each element in the expression represents the degree of contribution of muscle 1 to the synergy of muscle 2. This is the co-activation matrix, which defines the temporal pattern of muscle 1 co-activation and represents the contribution of co-activation 2 to the overall muscle activation pattern at a certain moment. e is the error between the reconstructed signal and the original signal. Here, k is the number of co-activations, and m and n are the input signal data of length n, composed of m channels of electromyography. The main control MCU performs non-negative matrix decomposition of the five channels of electromyography signals and directly outputs four co-activations. This results in less data transmission, higher efficiency, reduced system power consumption, and is beneficial for the long-term continuous operation of wearable devices.

[0039] After receiving the data processed by NNMF and the acceleration data, the host computer obtains the COP (center of pressure) and GRF (ground reaction force) based on the above data, and determines the user's status based on the obtained COP and GRF.

[0040] In the embodiments of this application, the host computer obtains COP and GRF through a pre-trained prediction model. For example... Figure 2As shown, the prediction model includes: an input layer, an LSTM layer, a COH collaborative consistency layer, an acceleration feature layer, a feature fusion layer, and a fully connected output layer.

[0041] Specifically, normalized multidimensional surface electromyography (SEMG) time-series data is input to the input layer. The SEMG time-series data is then input to an LSTM layer to obtain LSTM features. The SEMG time-series data is then input to a CoH (Coherence to Harmony) coherence layer to obtain CoH (Coherence to Harmony) features. An acceleration feature layer is used to extract acceleration features from the acceleration data. A feature fusion layer fuses the LSTM features, CoH features, and acceleration features to obtain fused features. The fused features are then input to a fully connected output layer for regression to obtain two output time series: predicted COP (Coherence to Harmony) and GRF (Gross Response Rate). Preferably, the LSTM layer is a BiLSTM layer, and the SEMG time-series data is input to the BiLSTM layer to obtain BiLSTM features. The feature fusion layer fuses the BiLSTM features, CoH features, and acceleration features to obtain fused features.

[0042] In a preferred embodiment, a normalized multidimensional time series of size 54*k*n is first fed into the input layer. Here, 54 represents 54 time segments of low, medium, and high walking speeds for multiple users, k represents the number of muscle coordinations, and n represents the length of the time series.

[0043] Furthermore, the multidimensional time series is input into the BILSTM layer and the COH layer respectively. The LSTM network is an extension of the RNN model, which includes memory units based on forget gates, input gates, and output gates. The forget gate is responsible for deciding whether to retain or delete existing information, the input gate determines the degree to which new information is added to the memory unit, and the output gate controls the contribution of the current value in the memory unit to the output. Due to its unique memory unit structure that considers the relationship between the preceding and following time series, the LSTM layer is considered to effectively solve the gradient vanishing problem and is suitable for predicting long time series. Furthermore, the BILSTM layer, based on the LSTM layer, can simultaneously consider both forward and backward time series information. The BILSTM layer runs one forward and one backward LSTM simultaneously, processing the forward and backward input sequences at each time step, thereby obtaining more complete feature information. Specifically, three cascaded BILSTM layers are used, each BILSTM unit including 512 hidden layers, and the obtained BiLSTM features are fed into a dropout layer with a probability of 0.3.

[0044] The COH coherence layer calculates wavelet correlations among multiple input time series within a certain frequency range to obtain wavelet correlation features representing the spatial correlation between them. The correlation features of multiple time series are crucial for multi-time series prediction and classification. On the one hand, multi-time series correlation can be used to filter redundant features; on the other hand, it can supplement the spatial dimension of features, particularly characterizing local features of multiple time series. A series of studies have attempted to use wavelet domain features based on electromyography (EMG) signals to represent the spatial features of multiple time series. However, a challenge of using EMG wavelet domain features for time series prediction and classification is the high feature dimensionality, which can lead to inaccurate predictions due to overfitting and other factors. In this application, muscle co-activation time series are used as the input time segment, which exhibits sparsity compared to wavelet domain features of EMG signals. Furthermore, muscle coherence features are extracted within the 4-40Hz range to further reduce sparsity while avoiding overfitting.

[0045] The COH collaborative consistency layer calculates the wavelet correlation features of the input multi-time series, and outputs a dimension of [dimensionality missing] for each time segment. The multidimensional features of. The feature dimension representing the wavelet correlation of multiple time series at each time step is denoted by n, where n represents the length of the multiple time series. Similarly, the COH features are fed into a dropout layer with a probability of 0.3.

[0046] In the embodiments of this application, Morlet wavelets are used as basis functions and a 1-second Hanning window is set to calculate the wavelet coefficients of multiple muscle co-activation time series within each time window. Morlet wavelets are complex-valued wavelet basis functions that exhibit good localization properties in both the time and frequency domains, making them suitable for analyzing the time-frequency characteristics of signals. and Represents co-activation sequence and In the The wavelet coefficients obtained by wavelet transform at each center frequency can be further used to calculate their cross power spectral density. for,

[0047]

[0048] Furthermore, the wavelet correlation is,

[0049]

[0050] in, and They represent the first time. The first frequency band and the The self-power spectrum of a co-activation sequence, assuming the number of co-activations meeting the requirements within the time window is k, the correlation between each pair of k co-activations will produce... Different combinations, therefore, the wavelet correlation of muscle coordination subscript Size from 1 to .

[0051] Finally, the BiLSTM features, COH features, and acceleration features are concatenated for feature fusion, and the concatenated features are input into a fully connected layer with 256 neurons. Further, the output feature information is passed through a dropout layer with a probability of 0.3 and a fully connected layer with 2 neurons output before being regressed to the COP and GRF output time series.

[0052] In one embodiment of the training process, the RMS / SE of the predicted values ​​of COP / GRF and the measured values ​​(which can be based on the pressure input signal of this invention) is used as the loss function. The Adam optimizer is employed, with an initial default learning rate of 0.0005, which is increased to 0.001 after 400 epochs. The optimal model is trained for 500 epochs; if the prediction performance does not improve for 50 consecutive epochs, the training is stopped early. Input data is fed into the model in batches of 512 to improve learning efficiency, and a shuffle is performed once per epoch.

[0053] In the embodiments of this application, two baseline methods, BILSTM and CNN-BILSTM, which respectively perform well in processing temporal and spatial local features of long-term multivariate time series, are used for comparative experiments. The BILSTM model shows good results and great potential in the continuous prediction of multivariate time series electrophysiological signals. A BILSTM model with three hidden layers (512 layers) and a dropout layer with a ratio of 0.3 is used. The CNN-BILSTM model concatenates a CNN model and a BILSTM model. First, the CNN model is used to extract deep features f, and then the deep features f from multiple time series are fed into the BILSTM model for time series regression estimation. The CNN model consists of one convolutional layer, one ReLU layer, one pooling layer, and one dropout layer. The CNN convolutional layer has 16 kernels with a kernel size of 3, the pooling layer has a size of 2*1, and the dropout layer has a ratio of 0.3. The learning rate of both baseline methods is 0.0005, which is increased to 0.001 after 400 epochs. The maximum training epoch is 500.

[0054] The processor, as the core computing unit of the device, is responsible for running applications, drivers, and the interaction logic between hardware and software.

[0055] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the present invention in any way, and all technical solutions obtained by equivalent substitution or equivalent transformation fall within the protection scope of the present invention.

Claims

1. A wearable human condition assessment device, characterized in that, Include: An electromyography (EMG) signal acquisition module is used to acquire the user's EMG signals, and the EMG signal acquisition module contains five channels; The MCU module is connected to the electromyography signal acquisition module and is used to receive the electromyography signals acquired by the electromyography signal acquisition module and process the received electromyography signals. The IMU module is used to collect the user's acceleration data; A wireless transmission module is connected to the MCU module and the IMU module. The wireless transmission module receives the data processed by the MCU module and the acceleration data collected by the IMU module, and wirelessly transmits them to the host computer.

2. The wearable human condition assessment device according to claim 1, characterized in that, The wireless transmission module includes at least one of a WiFi module or a Bluetooth module.

3. The wearable human condition assessment device according to claim 1, characterized in that, The wireless transmission module is wirelessly connected to the IMU module.

4. The wearable human condition assessment device according to claim 1, characterized in that, The MCU module includes a computing unit and five ADC units. The five ADC units perform analog-to-digital conversion on the five channels of electromyography signals, and the computing unit performs NNMF processing on the converted multi-channel electromyography signals.

5. The wearable human condition assessment device according to claim 4, characterized in that, The host computer obtains COP and GRF based on the received NNMF-processed data and acceleration data, and determines the user's state based on the obtained COP and GRF.

6. The wearable human condition assessment device according to claim 5, characterized in that, The host computer obtains COP and GRF through a pre-trained prediction model.

7. The wearable human condition assessment device according to claim 6, characterized in that, The prediction model comprises: an input layer, an LSTM layer, a COH collaborative consistency layer, an acceleration feature layer, a feature fusion layer, and a fully connected output layer; Normalized multidimensional surface electromyography time series data are input into the input layer; Multidimensional surface electromyography time-series data are input into the LSTM layer to obtain LSTM features; Multidimensional surface electromyography time-series data are input into the COH coherence layer to obtain COH features; The acceleration feature layer is used to extract acceleration features from the acceleration data; The feature fusion layer fuses LSTM features, COH features, and acceleration features to obtain fused features; The fused features are input into the fully connected output layer to obtain the predicted COP and GRF.

8. The wearable human condition assessment device according to claim 7, characterized in that, The LSTM layer is a BiLSTM layer; Multidimensional surface electromyography time-series data are input into the BiLSTM layer to obtain BiLSTM features; The feature fusion layer fuses BiLSTM features, COH features, and acceleration features to obtain fused features.

9. The wearable human condition assessment device according to claim 8, characterized in that, The BILSTM layer uses either the BILSTM model or the CNN-BILSTM model.

10. The wearable human condition assessment device according to claim 7, characterized in that, The Adam optimizer is used when training the prediction model.

Images