Motion mode recognition device and method for knee joint exoskeleton robot

Abstract

The invention discloses a motion mode recognition device and method for a knee joint exoskeleton robot, and the device comprises a lower limb motion measurement module which is used for measuring lower limb motion data; the waist dragging force measuring module is used for measuring the dragging force borne by the waist; the controller module is used for resolving and recognizing the lower limb movement data and the waist dragging force according to a deep neural network movement mode recognition model and determining a current movement mode; calculating to obtain the average tension borne by the waist; the result output module is used for outputting the current motion mode and the average tension borne by the waist in a voice broadcast mode; and the fixing device is used for fixing the motion mode recognition device and the knee joint exoskeleton robot. According to the scheme, the waist dragging force and the inertial sensor data are fused to serve as input of a motion mode recognition algorithm, the waist dragging force is innovatively converted and applied to the algorithm, the data input dimension is expanded, and the recognition accuracy is effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of exoskeleton motion pattern recognition technology, and particularly relates to a motion pattern recognition device and method for a knee joint exoskeleton robot. Background Technology

[0002] Exoskeleton robots are intelligent systems that conform to human anatomy and can be worn outside the human body. They sense human movement and synchronously "follow" it, using energy and mechanical power to enhance human strength, speed, and endurance. They represent one of the most deeply human-machine integrated pieces of equipment in the field of robotics. Exoskeleton robots require different gait control strategies for different individuals and environments; failure to adjust gait strategies in a timely manner hinders stable and efficient control.

[0003] Human movement is characterized by high autonomy, high degree of freedom, complex information, and diverse actions. Currently, there are three main methods for realizing human movement patterns: The first method collects bioelectrical signals through biomolecular electromyography (EMG) sensors and EEG sensors to achieve motion perception and decision-making. This method has the following drawbacks: it requires a lot of time for sensor deployment, is susceptible to interference from collisions between humans and machines, has insufficient anti-interference capabilities, and the wearing process is relatively cumbersome. The second method is based on visual sensors for recognition. This method has the following drawbacks: unstable camera field of view and jitter can lead to unstable capture of the surrounding environment, and it is susceptible to ambient light interference. The third method is based on inertial sensors for recognition. This method has the following drawbacks: the transmission distance of wireless communication methods limits the ability to achieve real-time recognition of movement patterns in outdoor environments. In addition, the design of knee-jointed exoskeleton robots generally neglects the stress on the lower back. Summary of the Invention

[0004] The technical problem solved by the present invention is to overcome the shortcomings of the prior art and provide a motion pattern recognition device and method for a knee exoskeleton robot. Based on data from multiple sensors, the device can classify motion patterns and assess the pulling force exerted on the human waist by the knee exoskeleton robot by collecting the drag force on the waist. The result output module provides the current motion pattern and the average pulling force on the waist.

[0005] To address the aforementioned technical problems, this invention discloses a motion pattern recognition device for a knee joint exoskeleton robot, comprising: The lower limb motion measurement module is used to measure and obtain lower limb motion data; The waist drag force measurement module is used to measure the drag force experienced by the waist. The controller module is used to solve and identify lower limb motion data and waist drag force based on a deep neural network motion pattern recognition model to determine the current motion pattern; and to calculate the average tension on the waist based on the drag force on the waist. The results output module is used to output the current motion mode and the average tension on the waist via voice broadcast. A fixation device is used to secure the motion pattern recognition device to the knee exoskeleton robot.

[0006] In the motion pattern recognition device of the aforementioned knee exoskeleton robot, the lower limb motion measurement module includes: inertial sensor A, inertial sensor B, inertial sensor C, and inertial sensor D; wherein, inertial sensor A and inertial sensor B are respectively installed on the outer side of the left and right lower legs of the knee exoskeleton robot; inertial sensor C and inertial sensor D are respectively installed on the outer side of the left and right thighs of the knee exoskeleton robot; the X-axis of each inertial sensor is vertically installed with the positive direction of the X-axis pointing upward, that is, the XZ axis is parallel to the sagittal plane of the body; the Y-axis of each inertial sensor is inward; each inertial sensor includes a lower limb CAN communication transceiver module, through which the inertial sensor measurement data is sent to the controller module.

[0007] In the motion pattern recognition device of the aforementioned knee exoskeleton robot, inertial sensor A is used to measure the motion data of the left lower leg, including: the angular velocity and angle of the left lower leg rotating around the Y-axis, the angle and angular velocity around the Z-axis, and the triaxial acceleration; inertial sensor B is used to measure the motion data of the right lower leg, including: the angular velocity and angle of the right lower leg rotating around the Y-axis, the angle and angular velocity around the Z-axis, and the triaxial acceleration; inertial sensor C is used to measure the motion data of the left thigh, including: the angular velocity and angle of the left thigh rotating around the Y-axis, the angle and angular velocity around the Z-axis, and the triaxial acceleration; and inertial sensor D is used to measure the motion data of the right thigh, including: the angular velocity and angle of the right thigh rotating around the Y-axis, the angle and angular velocity around the Z-axis, and the triaxial acceleration.

[0008] In the motion pattern recognition device of the above-mentioned knee exoskeleton robot, the waist drag force measurement module includes: elastic tension module E and elastic tension module F; wherein, elastic tension module E is installed on the outer side of the left hip of the knee exoskeleton robot to measure the drag force of the left waist; elastic tension module F is installed on the outer side of the right hip of the knee exoskeleton robot to measure the drag force of the right waist.

[0009] In the motion pattern recognition device of the aforementioned knee exoskeleton robot, the elastic tension module E and the elastic tension module F have the same structure, including: an elastic strain sensor, a tension acquisition plate, and a waist CAN communication transceiver module; wherein, the elastic strain sensor is based on the principle of capacitive strain sensing, and deformation causes a change in capacitance value; the tension acquisition plate converts the capacitance signal output by the elastic strain sensor into tension, thereby obtaining the drag force on the waist, and sends it to the controller module through the waist CAN communication transceiver module.

[0010] In the motion pattern recognition device of the aforementioned knee exoskeleton robot, the conversion from capacitive signal to tensile force is accomplished based on the following formula:

[0011] The elastic strain sensor consists of two plates. Indicates tension. Represents the vacuum permittivity. Represents the relative permittivity of the medium. Indicates the area of ​​the electrode plate. This represents the initial distance between the two plates when no external force is applied. This indicates the change in capacitance. This represents the excitation voltage.

[0012] In the motion pattern recognition device of the aforementioned knee exoskeleton robot, the controller module adopts an MCU+AISOC architecture design. The MCU is used to realize data synchronization and filtering, and sends the preprocessed data to the AISOC module through SPI communication. The AISOC module is used to receive the preprocessed data and perform calculations and recognition based on the deep neural network motion pattern recognition model to determine the current motion pattern and the average tension on the waist.

[0013] In the motion pattern recognition device of the aforementioned knee exoskeleton robot, when the controller module calculates and identifies lower limb motion data and lumbar drag force based on a deep neural network motion pattern recognition model to determine the current motion pattern, it includes: A deep neural network motion pattern recognition model is constructed based on the multi-scale dilated convolution and median fusion method. The received lower limb motion data and the drag force on the waist are filtered and normalized to obtain the input data x; The input data x is used as input to a deep neural network motion pattern recognition model to perform calculations, recognition, and determine the current motion pattern.

[0014] In the motion pattern recognition device of the aforementioned knee exoskeleton robot, the deep neural network motion pattern recognition model divides the calculation and recognition of input data x into the following four processing stages: The first processing stage employs five parallel convolutional branches to extract features from the input data x at different scales, ultimately outputting features y. Each convolutional branch is configured with a different dilation rate to expand the receptive field and capture spatial information of varying ranges. The first convolutional branch uses a 1x1 kernel, extracting features directly without altering the spatial scale. The second convolutional branch uses a 3x3 kernel with a dilation rate of S1 to expand the receptive field. The third convolutional branch uses a 3x3 kernel with a dilation rate of S2 to further expand the receptive field and capture a wider range of contextual information. The fourth convolutional branch uses a 3x3 kernel with a dilation rate of S3 to provide the widest receptive field. The fifth convolutional branch uses global average pooling to extract global contextual features, enhancing the model's understanding of the overall layout. The second processing stage consists of global average pooling (AvgPool), global max pooling (MaxPool), global median pooling (MedianPool), shared MLP fusion processing, and weighted features; after processing by the second processing stage, feature y outputs purified channel feature z. The third processing stage consists of the fusion of channel attention and spatial attention element addition operations; the purified channel feature z is processed by the third processing stage and outputs a spatiotemporal hybrid vector u; Fourth processing stage: The spatiotemporal hybrid vector u is output as a target vector v after passing through a fully connected network; the target vector v is a 6x1 feature vector, realizing the recognition of six movement modes: walking, running, climbing stairs, descending stairs, uphill, and downhill.

[0015] Accordingly, the present invention also discloses a motion pattern recognition method for a knee joint exoskeleton robot, comprising: A deep neural network motion pattern recognition model is constructed based on the multi-scale dilated convolution and median fusion method. The measurement yielded lower limb movement data and the drag force experienced by the waist; The measured lower limb motion data and the drag force on the waist are filtered and normalized to obtain input data; the input data is then processed and identified using a deep neural network motion pattern recognition model to determine the current motion pattern. Based on the measured drag force on the waist, the average tension on the waist is calculated.

[0016] The present invention has the following advantages: (1) This invention discloses a motion pattern recognition device for a knee joint exoskeleton robot, which integrates waist drag force and inertial sensor data as input to the motion pattern recognition algorithm. It innovatively transforms waist belt drag force and applies it to the algorithm, expands the data input dimension, and effectively improves the recognition accuracy.

[0017] (2) This invention discloses a motion pattern recognition device for a knee exoskeleton robot. Based on the output motion pattern, it adds a lumbar drag force assessment. It innovatively proposes to use the lumbar drag force to assess the pulling force applied to the waist by the knee exoskeleton robot. Compared with the use of pressure sensors, the hip belt can be easily integrated into the overall design of the knee exoskeleton robot. It is easy to use with the hip belt, and it is not easy for users to feel foreign objects. It is also easy to wear.

[0018] (3) This invention discloses a motion pattern recognition device for a knee joint exoskeleton robot. It innovatively proposes a motion pattern recognition module based on multi-scale dilated convolution and median fusion method. Through parallel dilation, local and global features at different scales are extracted from the network without expanding the receptive field. Combined with median enhancement technology, the feature extraction capability is further improved, and features can be captured and fused at different scales.

[0019] (4) This invention discloses a motion pattern recognition device for a knee exoskeleton robot. It innovatively proposes a controller architecture based on MCU+AISOC, which effectively realizes the embedded deployment of motion pattern recognition algorithm and meets the real-time and environmental complexity requirements of the knee exoskeleton robot. Attached Figure Description

[0020] Figure 1 This is a block diagram of a motion pattern recognition device for a knee exoskeleton robot according to an embodiment of the present invention; Figure 2 This is a schematic diagram (front view) of the layout of a motion pattern recognition device according to an embodiment of the present invention. Figure 3 This is a side view of the layout of a motion pattern recognition device according to an embodiment of the present invention. Figure 4 This is a structural schematic diagram of an elastic tension module according to an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the working principle of a deep neural network motion pattern recognition model in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments disclosed in the present invention will be described in further detail below with reference to the accompanying drawings.

[0022] Reference Figure 1In this embodiment, the motion pattern recognition device for the knee exoskeleton robot includes: a lower limb motion measurement module, a lumbar drag force measurement module, a controller module, and a result output module. The lower limb motion measurement module measures lower limb motion data; the lumbar drag force measurement module measures the drag force acting on the lumbar region; the controller module calculates and identifies the lower limb motion data and lumbar drag force based on a deep neural network motion pattern recognition model to determine the current motion pattern; and calculates the average tension on the lumbar region based on the drag force acting on the lumbar region; the result output module outputs the current motion pattern and the average tension on the lumbar region via voice broadcast; and a fixing device secures the motion pattern recognition device to the knee exoskeleton robot.

[0023] In this embodiment, as Figures 2-3 As shown, the lower limb motion measurement module mainly includes: inertial sensor A1, inertial sensor B2, inertial sensor C3, and inertial sensor D4. Inertial sensor A1 is installed on the outer side of the left lower leg of the knee exoskeleton robot to measure the motion data of the left lower leg (angular velocity and angle of rotation around the Y-axis, angle and angular velocity around the Z-axis, and triaxial acceleration). Inertial sensor B2 is installed on the outer side of the right lower leg of the knee exoskeleton robot to measure the motion data of the right lower leg (angular velocity and angle of rotation around the Y-axis, angle and angular velocity around the Z-axis, and triaxial acceleration). Inertial sensor C3 is installed on the outer side of the left thigh of the knee exoskeleton robot to measure the motion data of the left thigh (angular velocity and angle of rotation around the Y-axis, angle and angular velocity around the Z-axis, and triaxial acceleration). Inertial sensor D4 is installed on the outer side of the right thigh of the knee exoskeleton robot to measure the motion data of the right thigh (angular velocity and angle of rotation around the Y-axis, angle and angular velocity around the Z-axis, and triaxial acceleration). Each inertial sensor is mounted vertically with its X-axis pointing upwards, meaning its XZ-axis is parallel to the body's sagittal plane; the Y-axis of each inertial sensor points inwards; each inertial sensor includes a lower limb CAN communication transceiver module, through which inertial sensor measurement data is sent to the controller module.

[0024] In this embodiment, as Figures 2-3 As shown, the waist drag force measurement module mainly includes: elastic tension module E5 and elastic tension module F6. Elastic tension module E5 can be installed on the outer side of the left hip of the knee exoskeleton robot via the hip strap in the fixing device, and is used to measure the left waist drag force; elastic tension module F6 can be installed on the outer side of the right hip of the knee exoskeleton robot via the hip strap in the fixing device, and is used to measure the right waist drag force.

[0025] Furthermore, the elastic tensile module E5 and the elastic tensile module F6 have the same structure, including: an elastic strain sensor 61, a tensile acquisition board 62, and a waist-mounted CAN communication transceiver module. Among them, as... Figure 4 As shown, the elastic strain sensor 61 is installed on the hip belt. The elastic strain sensor 61 is based on the principle of capacitive strain sensing, and deformation causes a change in capacitance value. The tensile acquisition board 62 converts the capacitance signal output by the elastic strain sensor 61 into tensile force, thereby obtaining the drag force on the waist, and sends it to the controller module through the waist CAN communication transceiver module.

[0026] Furthermore, the conversion from capacitance signal to tension can be achieved based on the following formula:

[0027] The elastic strain sensor consists of two plates. Indicates tension. Represents the vacuum permittivity. Represents the relative permittivity of the medium. Indicates the area of ​​the electrode plate. This represents the initial distance between the two plates when no external force is applied. This indicates the change in capacitance. This represents the excitation voltage.

[0028] In this embodiment, the controller module adopts an MCU+AISOC architecture design. The MCU is used for data synchronization and filtering, and sends preprocessed data to the AISOC module via SPI communication. The AISOC module receives the preprocessed data and, based on a deep neural network motion pattern recognition model, performs calculations and identifications to determine the current motion pattern and the average tension on the waist.

[0029] In this embodiment, when the controller module calculates and identifies lower limb motion data and lumbar drag force based on a deep neural network motion pattern recognition model to determine the current motion pattern, it includes: (1) A deep neural network motion pattern recognition model was constructed based on the multi-scale dilated convolution and median fusion method. In order to extract local and global features of different scales from the network, the deep neural network motion pattern recognition model was constructed based on the multi-scale dilated convolution and median fusion method.

[0030] (2) The received lower limb motion data and the drag force on the waist are filtered and normalized to obtain the input data x.

[0031] Filtering: The received data (received lower limb motion data and drag force on the waist) is segmented using a sliding window, and then filtered using a Butterworth filter. The amplitude squared function of the Butterworth filter... It is expressed as follows:

[0032] in, This indicates that the Butterworth filter operates at angular frequency. The amplitude response at that point, Represents angular frequency. Indicates the cutoff frequency. Indicates the filter order.

[0033] Normalization: Normalization helps find the global optimum during later training; it maps the data to the [0,1] space. The normalization formula is as follows:

[0034] in, and These represent the maximum and minimum values ​​in the data within the current sampling period, respectively. Indicates any number of samples within the current sampling period Data.

[0035] (3) The input data x is used as the input to the deep neural network motion pattern recognition model for calculation and recognition to determine the current motion pattern, such as... Figure 5 As shown, the process is divided into the following four stages: The first processing stage employs five parallel convolutional branches to extract features from the input data x at different scales, ultimately outputting features y. Each convolutional branch has a different dilation rate to expand the receptive field and capture spatial information across different ranges. The first convolutional branch uses a 1x1 kernel, extracting features directly without altering the spatial scale. The second convolutional branch uses a 3x3 kernel with a dilation rate of S1 to expand the receptive field. The third convolutional branch uses a 3x3 kernel with a dilation rate of S2 to further expand the receptive field and capture broader contextual information. The fourth convolutional branch uses a 3x3 kernel with a dilation rate of S3 to provide the widest receptive field. The fifth convolutional branch uses global average pooling to extract global contextual features, enhancing the model's understanding of the overall layout.

[0036] The second processing stage consists of global average pooling (AvgPool), global max pooling (MaxPool), global median pooling (MedianPool), shared MLP fusion processing, and weighted features; after the second processing stage, the feature y is processed to output the purified channel feature z.

[0037] The third processing stage consists of the fusion of channel attention and spatial attention element addition operations; the purified channel features z are processed by the third processing stage and output as a spatiotemporal hybrid vector u.

[0038] Fourth processing stage: The spatiotemporal hybrid vector u is output as a target vector v after passing through a fully connected network; the target vector v is a 6x1 feature vector, realizing the recognition of six movement modes: walking, running, climbing stairs, descending stairs, uphill, and downhill.

[0039] In this embodiment, when the controller module calculates the average tension force on the waist based on the drag force acting on the waist, it can specifically calculate the average tension force on the waist using the following formula. :

[0040] in, The module indicating the waist drag force measurement The second measurement obtained the drag force on the waist. This represents one measurement cycle. The average tension applied to the waist... Greater than a certain threshold When the time is right, it indicates that there is a deviation in the human-machine binding, affecting the wearer's comfort, and the binding needs to be redone.

[0041] In this embodiment, the fixing device mainly includes: a waist belt 7 and a hip belt 8. The controller module 9 is installed on the waist of the knee exoskeleton robot through the waist belt 7; the elastic tension module is installed on the outer side of the left and right hips of the knee exoskeleton robot through the hip belt 8.

[0042] In summary, the motion pattern recognition device for the knee exoskeleton robot described in this invention recognizes motion patterns based on multi-sensor fusion, assisting the knee exoskeleton robot in identifying the wearer's motion patterns. Simultaneously, it quantifies the ergonomics of the lumbar region, allowing users to clearly understand the comfort level of the device. Using lumbar drag force as the motion pattern sensing input expands the data input dimensions and effectively improves recognition accuracy. After comprehensive field testing, the recognition accuracy increased from 96.67% to 98.25% compared to not incorporating lumbar drag force. Compared to using pressure sensors, the hip belt design can be easily integrated into the overall design of the knee exoskeleton robot, requiring no additional operation. Actual testing shows that when the lumbar drag force exceeds 5N, wearing comfort is poor, affecting the user experience of the knee exoskeleton robot. This motion pattern recognition model, based on multi-scale dilated convolution and median fusion, extracts local and global features at different scales from the network through parallel dilation without expanding the receptive field. Combined with median enhancement, it further improves feature extraction capabilities, enabling the capture and fusion of features at different scales. Compared to other methods (such as CNN-LSTM-Attention), this method achieves an accuracy of 97.63% on the public dataset HuGaDB, representing an accuracy improvement of over 2%. The MCU+AISOC controller architecture effectively enables embedded deployment of the motion pattern recognition model, meeting the real-time and environmental complexity requirements of knee exoskeleton robots. Actual testing shows that the single-run time of this architecture is <20ms, satisfying the real-time requirements of knee exoskeleton robot products. The output module provides voice prompts, providing real-time reminders to the wearer about the current motion mode and lumbar stress.

[0043] Based on the above embodiments, this invention also discloses a motion pattern recognition method for a knee exoskeleton robot, comprising: constructing a deep neural network motion pattern recognition model based on multi-scale dilated convolution and median fusion; measuring lower limb motion data and the drag force on the waist; filtering and normalizing the measured lower limb motion data and the drag force on the waist to obtain input data; solving and recognizing the input data through the deep neural network motion pattern recognition model to determine the current motion pattern; calculating the average tension on the waist based on the measured drag force on the waist; and solving and recognizing the lower leg motion data and waist tension data according to the deep neural network motion pattern recognition model to obtain the motion pattern and output it.

[0044] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications to the technical solutions of the present invention by utilizing the methods and techniques disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall fall within the protection scope of the technical solutions of the present invention.

[0045] The contents not described in detail in this specification are common knowledge to those skilled in the art.

Claims

1. A motion pattern recognition device for a knee exoskeleton robot, characterized by, include: The lower limb motion measurement module is used to measure and obtain lower limb motion data; The waist drag force measurement module is used to measure the drag force experienced by the waist. The controller module is used to solve and identify lower limb motion data and waist drag force based on a deep neural network motion pattern recognition model to determine the current motion pattern; and to calculate the average tension on the waist based on the drag force on the waist. The results output module is used to output the current motion mode and the average tension on the waist via voice broadcast. A fixation device is used to secure the motion pattern recognition device to the knee exoskeleton robot.

2. The motion pattern recognition device of the knee exoskeleton robot according to claim 1, characterized in that, The lower limb motion measurement module includes: inertial sensor A (1), inertial sensor B (2), inertial sensor C (3) and inertial sensor D (4); wherein, inertial sensor A (1) and inertial sensor B (2) are respectively installed on the outer side of the left lower leg and the outer side of the right lower leg of the knee joint exoskeleton robot; inertial sensor C (3) and inertial sensor D (4) are respectively installed on the outer side of the left thigh and the outer side of the right thigh of the knee joint exoskeleton robot; the X-axis of each inertial sensor is vertically installed with the positive direction of the X-axis pointing upward, that is, the XZ axis is parallel to the sagittal plane of the body; the Y-axis of each inertial sensor is inward; each inertial sensor contains a lower limb CAN communication transceiver module, through which the inertial sensor measurement data is sent to the controller module.

3. The motion pattern recognition device of the knee exoskeleton robot according to claim 2, wherein Inertial sensor A (1) is used to measure the motion data of the left lower leg, including: the angular velocity and angle of the left lower leg rotating around the Y-axis, the angle and angular velocity around the Z-axis, and the triaxial acceleration; Inertial sensor B (2) is used to measure the motion data of the right lower leg, including: the angular velocity and angle of the right lower leg rotating around the Y-axis, the angle and angular velocity around the Z-axis, and the triaxial acceleration; Inertial sensor C (3) is used to measure the motion data of the left thigh, including: the angular velocity and angle of the left thigh rotating around the Y-axis, the angle and angular velocity around the Z-axis, and the triaxial acceleration; Inertial sensor D (4) is used to measure the motion data of the right thigh, including: the angular velocity and angle of the right thigh rotating around the Y-axis, the angle and angular velocity around the Z-axis, and the triaxial acceleration.

4. The motion pattern recognition device of the knee exoskeleton robot according to claim 1, wherein The waist drag force measurement module includes: elastic tension module E (5) and elastic tension module F (6); wherein, elastic tension module E (5) is installed on the outer side of the left hip of the knee joint exoskeleton robot to measure the drag force of the left waist; elastic tension module F (6) is installed on the outer side of the right hip of the knee joint exoskeleton robot to measure the drag force of the right waist.

5. The motion pattern recognition device of the knee exoskeleton robot according to claim 4, wherein The elastic tension module E (5) and the elastic tension module F (6) have the same structure, including: elastic strain sensor (61), tension acquisition board (62) and waist CAN communication transceiver module; wherein, the elastic strain sensor (61) is based on the principle of capacitive strain sensing, and deformation causes the capacitance value to change; the tension acquisition board (62) converts the capacitance signal output by the elastic strain sensor (61) into tension, thereby obtaining the drag force on the waist, and sends it to the controller module through the waist CAN communication transceiver module.

6. The motion pattern recognition device of the knee exoskeleton robot according to claim 5, wherein The conversion from capacitance signal to tension is accomplished based on the following formula: Among them, the elastic strain sensor (61) consists of two plates; Indicates tension. Represents the vacuum permittivity. Represents the relative permittivity of the medium. Indicates the area of ​​the electrode plate. This represents the initial distance between the two plates when no external force is applied. This indicates the change in capacitance. This represents the excitation voltage.

7. The motion pattern recognition device for the knee joint exoskeleton robot according to claim 1, characterized in that, The controller module adopts an MCU+AISOC architecture design. The MCU is used to realize data synchronization and filtering, and sends the preprocessed data to the AISOC module through SPI communication. The AISOC module is used to receive the preprocessed data and perform calculations and recognition based on the deep neural network motion pattern recognition model to determine the current motion pattern and the average tension on the waist.

8. The motion pattern recognition device of the knee exoskeleton robot according to claim 7, wherein When the controller module calculates and identifies lower limb motion data and lumbar drag force based on a deep neural network motion pattern recognition model to determine the current motion pattern, it includes: A deep neural network motion pattern recognition model is constructed based on the multi-scale dilated convolution and median fusion method. The received lower limb motion data and the drag force on the waist are filtered and normalized to obtain the input data x; The input data x is used as input to a deep neural network motion pattern recognition model to perform calculations, recognition, and determine the current motion pattern.

9. The motion pattern recognition device of the knee exoskeleton robot according to claim 8, wherein The deep neural network motion pattern recognition model processes input data x in four stages: The first processing stage employs five parallel convolutional branches to extract features from the input data x at different scales, ultimately outputting features y. Each convolutional branch is configured with a different dilation rate to expand the receptive field and capture spatial information of varying ranges. The first convolutional branch uses a 1x1 kernel, extracting features directly without altering the spatial scale. The second convolutional branch uses a 3x3 kernel with a dilation rate of S1 to expand the receptive field. The third convolutional branch uses a 3x3 kernel with a dilation rate of S2 to further expand the receptive field and capture a wider range of contextual information. The fourth convolutional branch uses a 3x3 kernel with a dilation rate of S3 to provide the widest receptive field. The fifth convolutional branch uses global average pooling to extract global contextual features, enhancing the model's understanding of the overall layout. The second processing stage consists of global average pooling (AvgPool), global max pooling (MaxPool), global median pooling (MedianPool), shared MLP fusion processing, and weighted features; after processing by the second processing stage, feature y outputs purified channel feature z. The third processing stage consists of the fusion of channel attention and spatial attention element addition operations; the purified channel feature z is processed by the third processing stage and outputs a spatiotemporal hybrid vector u; Fourth processing stage: The spatiotemporal hybrid vector u is output as a target vector v after passing through a fully connected network; the target vector v is a 6x1 feature vector, realizing the recognition of six movement modes: walking, running, climbing stairs, descending stairs, uphill, and downhill.

10. A method for recognizing a motion pattern of a knee exoskeleton robot, characterized by, include: A deep neural network motion pattern recognition model is constructed based on the multi-scale dilated convolution and median fusion method. The measurement yielded lower limb movement data and the drag force experienced by the waist; The measured lower limb motion data and the drag force on the waist are filtered and normalized to obtain input data; the input data is then processed and identified using a deep neural network motion pattern recognition model to determine the current motion pattern. Based on the measured drag force on the waist, the average tension on the waist is calculated.

Images