Knee joint exoskeleton multi-action torque auxiliary control method based on activity classification and phase detection

Abstract

The invention provides a knee joint exoskeleton multi-action torque auxiliary control method based on activity classification and phase detection. The knee joint exoskeleton multi-action torque auxiliary control method comprises a two-stage artificial neural network and three human body dynamic models. The first stage of the two-stage artificial neural network is composed of a network used for activity classification, the second stage of the two-stage artificial neural network is composed of three independent gait phase estimation regression networks, and the three human body dynamic models comprise a single-support human body model, a double-support human body model and a swing human body model which are divided according to human body gait phase features. The model outputs the real biological joint torque at the current moment, and the controller proportionally outputs an auxiliary torque curve required by the exoskeleton. According to the method, four activity types of standing, walking, going upstairs and going downstairs can be accurately classified, the gait phase percentage under the non-rhythm gait condition is accurately estimated, and meanwhile, the corresponding auxiliary torque curve is generated based on the current gait type and gait phase.

Description

Technical Field

[0001] The present invention relates to the field of intelligent exoskeleton control, and in particular to a knee joint exoskeleton multi-action torque auxiliary control method based on activity classification and phase detection. Background Art

[0002] As a wearable robot, the lower limb exoskeleton system can work in conjunction with the human musculoskeletal system and has long been expected to enhance the user's athletic ability. The application of knee exoskeleton robots in the field of smart elderly care can specifically assist middle-aged and elderly people to improve their quality of life in community-based elderly care. It assists by applying extension and flexion torque to the knee joint, reducing the output of muscles such as the quadriceps, semitendinosus and soleus during exercise, and improving the wearer's athletic ability and safety. However, in daily life, efficiently assisting middle-aged and elderly people with knee arthritis or muscle weakness to perform daily activities is one of the most challenging and potential application goals of knee exoskeletons.

[0003] The ability to accurately and in real time identify a variety of non-rhythmic human motion intentions is the key to a high-performance control strategy for knee exoskeleton robots. Since exoskeleton robots face the problems of real-time scene changes and multi-dimensional continuous conversion of movements when used outdoors, traditional motion detection and gait estimation strategies originally suitable for laboratory testing or clinical rehabilitation applications have the risk of causing harm to the wearer. Therefore, it is extremely necessary to study the torque-assisted control method of knee exoskeleton based on community multi-motion applications.

[0004] Conventional active exoskeletons’ torque-assisted control strategies based on time (gait cycle) or kinematics (such as joint angles) cannot adapt well to changes in outdoor walking patterns. These controllers usually generate discrete auxiliary torque trajectories based on gait phases. Their auxiliary trajectories can only change at the beginning of the next discrete segment, making it difficult to adapt to different walking patterns (such as speed changes and sudden stops). Continuous gait detection and continuous torque control strategies are expected to solve the above problems. The auxiliary trajectories can be updated instantly according to input signals (such as joint angles), thereby quickly adapting to changes in walking patterns. Summary of the invention

[0005] The purpose of the present invention is to provide a knee exoskeleton multi-action torque assisted control method based on activity classification and phase detection, which can accurately classify four activity types: standing, walking, climbing stairs and descending stairs, and accurately estimate the gait phase percentage under non-rhythmic gait conditions, and generate a corresponding auxiliary torque curve based on the current gait category and gait phase.

[0006] To achieve the above object, the present invention provides a knee exoskeleton multi-action torque assisted control method based on activity classification and phase detection, comprising the following steps:

[0007] The kinematic data of both sides of the thigh, both sides of the calf and the chest are collected as the input vector X of the artificial neural network and the human body dynamics model;

[0008] Artificial neural network uses sliding time window method to extract features to achieve action classification;

[0009] The actual ground gait phase percentage used for model training is obtained by installing a foot switch on the sole of the foot;

[0010] Establish an activity classification neural network that takes the kinematic data of both sides of the thigh as input and outputs the current activity type label;

[0011] A gait phase detection neural network is established, with the kinematic data of both sides of the thigh as input. The gait phase detection network corresponding to the activity type is selected according to the activity type label output by the activity classification neural network, and the output is the current real-time gait phase.

[0012] Evaluate activity classification neural networks and gait phase detection neural networks;

[0013] A human body dynamics model is created. The human body dynamics model generates the real joint biological torque at the current moment according to the current gait phase, and the controller provides an auxiliary torque proportional to the real torque.

[0014] According to different gait phase modes, the gait phase is divided into single support phase, double support phase and swing phase, and corresponds to three different human body dynamics models. The kinematic data of both sides of the thigh, both sides of the calf and the chest are used as the input of the model to obtain the real biological joint torque at the current moment.

[0015] Preferably, the kinematic data includes joint angles, angular velocities and accelerations.

[0016] Preferably, the actual ground gait phase percentage is obtained by installing a foot switch on the sole of the foot, including

[0017] Each time the foot switch detected ground contact, the gait phase percentage was reset to 0%, the gait phase at the previous time point was set to 100%, and the data points between 0% and 100% were uniformly incremented to obtain the true gait phase percentage at each time point;

[0018] In the case where the actual duration is shorter than the estimated time, the gait phase percentage is reset to 0 before reaching 100%; if the actual duration is longer than the estimated time, the gait phase percentage remains at 100% until the next ground contact is detected.

[0019] Preferably, the activity classification neural network uses a 4×1 vector as output, and outputs the result of the classifier as a 4×1 binary vector, corresponding to the four activity types of standing, walking, climbing stairs and descending stairs, and the value of the corresponding activity type is set to 1, and the other activity types are 0.

[0020] Preferably, the gait phase detection neural network includes three independent networks, corresponding to three specific typical activity types: walking, climbing stairs and descending stairs.

[0021] Preferably, the current real-time gait phase includes:

[0022] The timestamp between two foot strikes is assigned a gait phase percentage in the range of 0-100%. Since the gait phase percentage jumps from 100% to 0% when the foot strikes the ground, the output one-dimensional result is mapped to a two-dimensional coordinate system by mapping the gait phase percentage to a two-dimensional coordinate (x, y); the two neurons in the output layer are used to output the two-dimensional mapping result of the gait phase percentage;

[0023] The current gait phase percentage is φ, and the mapping formula is:

[0024] x=cos(2πφ), y=sin(2πφ) (1);

[0025] Among them, x represents the abscissa of the two-dimensional coordinate system, and y represents the ordinate of the two-dimensional coordinate system;

[0026] The calculated two-dimensional mapping (x, y) is calculated by the following equation to obtain the estimated gait phase percentage φ at the current moment current :

[0027]

[0028] Preferably, the activity classification neural network and the gait phase detection neural network are evaluated, including

[0029] (1) Evaluation of activity classification neural networks

[0030] Use the timestamp to describe the precision r a and the detection delay Δt of the activity transition t To evaluate the activity classification network, the formula is as follows:

[0031] r a =T c / T (3);

[0032]

[0033] Among them, T cIndicates the number of correctly classified timestamps, T is the total number of timestamps, N t represents the total number of activity transitions, t t,n and Represents the detected value and true value of the nth activity transition time;

[0034] (2) Evaluation of gait phase detection neural network

[0035] The evaluation index of gait phase percentage detection is the relative root mean square error rRMSE of the gait phase percentage, and its specific implementation formula is as follows:

[0036]

[0037] in, is the gait phase percentage calculated by (2), φ t is the true value of the gait phase percentage at time t, Represents the mean of the gait phase percentages.

[0038] Preferably, the gait phase is divided into a single support phase, a double support phase and a swing phase, and the dynamic equations corresponding to the three different human body dynamics models are as follows:

[0039] Single-support human body model: The human body only needs one foot to touch the ground. By setting the weight and inertia parameters of the upper body, thigh and calf, the corresponding inertia matrix Dsp(θ)∈1×5, centrifugal force and Coriolis matrix in the single-support model are derived. And the gravity matrix Gsp(θ)∈1×5, the knee joint torque is derived by the following formula (6):

[0040]

[0041] Double support human body model: The human body needs to touch the ground with both feet. The knee joint torque is derived from formula (7). The corresponding inertia matrix in the double support model is Ddp(θ)∈1×5, the centrifugal force and Coriolis matrix The gravity matrix is ​​Gdp(θ)∈1×5. The knee joint moment is derived by the following formula (7):

[0042]

[0043] Swinging human body model: The legs are in the swing phase, and the knee joint torque is derived from formula (8). The corresponding inertia matrix in the swinging model is Ds(θ)∈1×5, the centrifugal force and the Coriolis matrix The gravity matrix is ​​Gs(θ)∈1×5, and the knee joint torque is derived by the following formula (8):

[0044]

[0045] A multi-action torque auxiliary control system for a knee joint exoskeleton based on activity classification and phase detection, comprising an artificial neural network and a human body dynamics model, wherein the artificial neural network is divided into a first-stage neural network and a second-stage neural network;

[0046] The first-stage neural network consists of an activity classification neural network, and the second-stage neural network consists of three independent gait phase detection neural networks. The activity classification neural network and the gait phase detection neural network are composed of four layers, namely, an input layer, two hidden layers, and an output layer. The first layer is the input layer, which includes 84 neurons; the second hidden layer is the encoder layer, which includes 20 neurons; the third hidden layer is the decoder layer, which includes 40 neurons.

[0047] The fourth output layer of the activity classification neural network contains 4 neurons, which represent the four activity types of standing, walking, climbing stairs and descending stairs respectively; the first three layers of the gait phase detection neural network have the same structure as the activity classification neural network, and the 2 neurons in the fourth output layer are used to output the gait phase at the current moment.

[0048] A terminal device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the steps of the method are implemented when the processor executes the computer program.

[0049] Therefore, the present invention adopts the above-mentioned knee joint exoskeleton multi-action torque auxiliary control method based on activity classification and phase detection, and the technical effects are as follows:

[0050] (1) This method only requires two IMUs to achieve accurate activity classification and gait phase detection;

[0051] (2) This method can accurately estimate the gait phase percentage in non-rhythmic gait;

[0052] (3) This method can switch the activity type in real time and respond quickly according to the actual situation;

[0053] (4) This method can control the exoskeleton to continuously assist gait activities in complex and changing terrains;

[0054] (5) This method can be dynamically modified to adapt to the gait specificity of different wearers;

[0055] (6) This method is almost completely applicable to all types of lower limb assistive devices currently available. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a flow chart of the multi-action torque-assisted control method of knee exoskeleton based on activity classification and phase detection;

[0057] Figure 2 Schematic diagram of a two-stage neural network architecture for activity classification and gait phase detection;

[0058] Figure 3 Schematic diagram of the foot switch used;

[0059] Figure 4 This is a schematic diagram of the neural network structure for activity classification;

[0060] Figure 5 Schematic diagram of the neural network structure for gait phase detection;

[0061] Figure 6 To use Figure 3 Schematic diagram of the results of the neural network identifying the activity type;

[0062] Figure 7 To use Figure 4 Schematic diagram of the comparison between the gait phase percentage detected by the neural network and the true phase percentage;

[0063] Figure 8 Schematic diagram of the auxiliary torque provided in proportion to the biological torque obtained by the multi-action torque auxiliary control method. DETAILED DESCRIPTION

[0064] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0065] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0066] Embodiment 1

[0067] A multi-action torque-assisted control system for knee exoskeleton based on activity classification and phase detection includes an artificial neural network part and a human dynamics model part. The entire neural network is divided into two stages: activity classification and gait phase detection. The specific network architecture is as follows: Figure 1As shown. The first-stage neural network consists of an activity classification neural network, and the second-stage neural network consists of three independent gait phase detection neural networks, each of which consists of four layers: one input layer, two hidden layers, and one output layer. The first layer is the input layer, which includes 84 neurons; the second hidden layer is the encoder layer, which includes 20 neurons; the third hidden layer is the decoder layer, which includes 40 neurons; the fourth output layer of the activity classification neural network contains 4 neurons, which represent the four activity types of standing, walking, climbing stairs, and going down stairs; the first three layers of the gait phase detection neural network have the same structure as the activity classification neural network, and the two neurons in the fourth output layer are used to output the gait phase at the current moment. The human body dynamics model generates the real joint biological torque at the current moment according to the current gait phase. According to different gait phase modes, the gait phase is divided into single support phase, double support phase and swing phase, corresponding to three different human body dynamics models. The angles, angular velocities and angular accelerations collected by the five IMUs are used as the input of the model to obtain the real biological joint torque at the current moment, and the controller provides an auxiliary torque proportional to the real torque.

[0068] like Figure 1 and Figure 2 As shown, the present invention adopts the above-mentioned knee joint exoskeleton multi-action torque auxiliary control method based on activity classification and phase detection, comprising the following steps:

[0069] Step 1: collect kinematic data of both sides of the thigh, both sides of the calf and the chest as the input vector X of the artificial neural network and the human body dynamics model;

[0070] Kinematic data include joint angles, angular velocities, and accelerations.

[0071] Five IMU sensors are installed 2 cm above the knee joints on the front of the thighs on both sides, 3 cm below the knee joints on the front of the calves on both sides, and in the depressions between the chest and the abdomen. The IMUs of the thighs and calves can monitor the joint angles, angular velocities, and accelerations, providing complete motion data, while the IMUs of the chest help capture the overall posture and center of gravity changes of the body, so that the model can better understand the force distribution and stability during movement. The sampling frequency of the IMU is 200Hz. Since the Euler angle data drifts over time during walking, the 3-axis acceleration and 3-axis angular velocity data of the sensor are finally determined as the input vectors of the artificial neural network and the human body dynamics model.

[0072] Step 2: The artificial neural network uses a sliding time window method to extract features to achieve action classification;

[0073] By extracting features within a fixed time period, local changes and contextual information of the data can be better captured, the impact of instantaneous noise can be reduced, and the stability of the features can be improved. In order to minimize the computational cost of the neural network and reduce the complexity of hardware and software, our experiments have proved that only the information of two IMU sensors on both thighs can accurately realize action classification while reducing the dimension of the feature vector. A 100ms sliding time window is used to extract features, and 7 features are collected for each signal channel during the sliding time window, including the maximum value, minimum value, average value, standard deviation, first data, median data and last data during the window period, so as to effectively capture the overall trend and dynamic changes of the time series and enhance the recognition effect of complex patterns.

[0074] Step 3: Get the actual ground gait phase percentage by installing a foot switch on the sole of the foot;

[0075] To calculate the loss function, an event-based method is used as the baseline method. During the training process, a tester is equipped with a Figure 3 The foot switch shown, each time the foot switch detects ground contact, the gait phase percentage is reset to 0%, the gait phase at the previous time point is set to 100%, and the data points between 0% and 100% are uniformly incremented to obtain the true gait phase percentage at each time point;

[0076] In the case where the actual duration is shorter than the estimated time, the gait phase percentage is reset to 0 before reaching 100%; if the actual duration is longer than the estimated time, the gait phase percentage remains at 100% until the next ground contact is detected.

[0077] Step 4: Establish an activity classification neural network, take the kinematic data of both sides of the thigh as input, and output the current activity type label;

[0078] The hidden layer uses the hyperbolic tangent activation function, the output layer uses the soft-max activation function, and the argmax function is used to obtain the activity type with the highest probability after calculation as the output result. The transmission process of the activity classification neural network input data is as follows:

[0079] E A =tanh(W e,A X+b e,A )(1);

[0080] D A =tanh(W d,A E A +b d,A )(2);

[0081] P A=tanh(W o,A D A +b o,A )(3);

[0082] activity = argmax(P A )(4);

[0083] Among them, X is the input vector of the entire network, E A is the encoding vector from the encoder layer, W e,A is the weight matrix of the encoder layer, b e,A is the encoder layer bias; D A is the vector from the decoder layer, W d,A is the weight matrix of the decoder layer, b d,A is the decoder layer bias; P A is a 4×1 vector representing the probability of each activity, W o,A is the output layer weight matrix, b o,A is the output layer bias. activity outputs the calculated current activity type label. If the calculated activity is standing, the output gait phase percentage is 0%. When the output result is other activity types, this activity type will be used to select the gait phase detection neural network corresponding to the current activity type in the second stage neural network.

[0084] The activity classification neural network uses a 4×1 vector as output and outputs the result of the classifier as a 4×1 binary vector, which corresponds to the four activity types of standing, walking, climbing stairs and descending stairs. The value of the corresponding activity type is set to 1, and the other activity types are 0. For example, the vector of walking is

[0100] . During the training process of the activity classification neural network, the minimization of the cross entropy function is used as its loss function. The specific implementation formula is as follows:

[0085]

[0086] Among them, y i is the binary label from the ground truth, p i The probability calculated for the ith activity of the network.

[0087] Step 5: Establish a gait phase detection neural network, take the kinematic data of both sides of the thigh as input, select the gait phase detection network corresponding to the activity type according to the activity type label output by the activity classification neural network, and output the current real-time gait phase;

[0088] The timestamp between two foot strikes is assigned a gait phase percentage in the range of 0-100%. Since the gait phase percentage jumps from 100% to 0% when the foot strikes the ground, in order to ensure the continuity of the detection results, the output one-dimensional results are mapped to a two-dimensional coordinate system, and the gait phase percentage is mapped to a two-dimensional coordinate (x, y) to maintain the continuity of the detection results. The two neurons in the output layer are used to output the two-dimensional mapping results of the gait phase percentage.

[0089] Assuming the current gait phase percentage is φ, the mapping formula is:

[0090] x=cos(2πφ), y=sin(2πφ)(6);

[0091] Among them, x represents the horizontal coordinate of the two-dimensional coordinate system, and y represents the vertical coordinate of the two-dimensional coordinate system.

[0092] The second stage of the gait phase detection neural network includes three independent networks, corresponding to three specific typical activity types: walking, climbing stairs, and descending stairs. The transmission process of the gait phase detection neural network is as follows:

[0093] E a =tanh(W e,a X+b e,a )(7);

[0094] D a =tanh(W d,a E a +b d,a ) (8);

[0095] [x,y] T =W o,a D a +b o,a (9);

[0096] Among them, E a is the encoding vector from the encoder layer, W e,a is the weight matrix of the encoder layer, b e,a is the encoder layer bias; D a is the vector from the decoder layer, W d,a is the weight matrix of the decoder layer, b d,a is the decoder layer bias; W o,a is the output layer weight matrix, b o,a is the bias for the output layer.

[0097] The calculated two-dimensional mapping (x, y) is calculated by the following equation to obtain the estimated gait phase percentage φ at the current moment current :

[0098]

[0099] The loss function of the second-stage gait phase detection neural network is:

[0100]

[0101] in, Represents the output of the neural network.

[0102] Step 6: Evaluate the activity classification neural network and the gait phase detection neural network;

[0103] Use the timestamp to describe the precision r a and the detection delay Δt of the activity transition t To evaluate the activity classification network, the formula is as follows:

[0104] r a =T c / T(12);

[0105]

[0106] Among them, T c Indicates the number of correctly classified timestamps, T is the total number of timestamps, N t represents the total number of activity transitions, t t,n and Represents the detected value and true value of the nth activity transition time. The activity classification results obtained by the activity classification neural network are as follows Figure 6 shown.

[0107] The evaluation index of gait phase percentage detection is the relative root mean square error rRMSE of the gait phase percentage, and its specific implementation formula is as follows:

[0108]

[0109] in, is the gait phase percentage calculated by (10), φ t is the true value of the gait phase percentage at time t, The comparison results between the detected gait phase and the real phase obtained by the gait phase detection neural network are shown in Figure 6 shown.

[0110] Step 7: Create a human body dynamics model. The human body dynamics model generates the real joint biological torque at the current moment according to the current gait phase and kinematic data, and the controller provides an auxiliary torque proportional to the real torque.

[0111] According to different gait phase modes, the gait phase is divided into single support phase, double support phase and swing phase, which correspond to the dynamic equations of three different human body dynamic models (single support model, double support model and swing model). A low-pass filter is applied to smooth the calculated torque trajectory. The kinematic data of both sides of the thigh, both sides of the calf and the chest are used as the input of the model to obtain the real biological joint torque at the current moment.

[0112] In different action categories, different percentage positions are selected to switch gait modes, and the switching position can be dynamically modified according to the true value of the foot switch to adapt to the gait specificity of different wearers. The input of the human dynamics model is set as the angle θ∈5×1 of the 5 IMUs, the angular velocity Angular acceleration The output is set as the knee joint torque τ joint ∈1×1. The dynamic equation is as follows:

[0113] Single-support human body model: The human body only needs one foot to touch the ground. By setting the weight and inertia parameters of the upper body, thigh and calf, the corresponding inertia matrix Dsp(θ)∈1×5, centrifugal force and Coriolis matrix in the single-support model are derived. And the gravity matrix Gsp(θ)∈1×5, the knee joint torque is derived by the following formula (15):

[0114]

[0115] Double support human body model: The human body needs to touch the ground with both feet. The knee joint torque is derived from formula (16). The corresponding inertia matrix in the double support model is Ddp(θ)∈1×5, the centrifugal force and Coriolis matrix The gravity matrix is ​​Gdp(θ)∈1×5. The knee joint moment is derived by the following formula (7):

[0116]

[0117] Swinging human body model: The legs are in the swing phase, and the knee joint torque is derived from formula (17). The corresponding inertia matrix in the swinging model is Ds(θ)∈1×5, the centrifugal force and the Coriolis matrix The gravity matrix is ​​Gs(θ)∈1×5, and the knee joint torque is derived by the following formula (8):

[0118]

[0119] The joint torque obtained is multiplied by the corresponding ratio to be used as the auxiliary torque of the knee exoskeleton. The specific ratio needs to be dynamically adjusted according to different wearers and different application scenarios. The auxiliary torque curve provided in proportion to the biological torque obtained by the multi-action torque auxiliary control method is as follows: Figure 8 shown.

[0120] A terminal device is provided in one embodiment of the present invention. The terminal device of this embodiment includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps in the above-mentioned method embodiments are implemented. Alternatively, when the processor executes the computer program, the functions of the modules / units in the above-mentioned device embodiments are implemented.

[0121] The computer program may be divided into one or more modules / units, and the one or more modules / units are stored in the memory and executed by the processor to accomplish the present invention.

[0122] The terminal device may be a computing device such as a desktop computer, a notebook, a PDA, a cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0123] The processor may be a central processing unit (CPU), a graphics processing unit (GPU), or other general-purpose processors.

[0124] The memory may be used to store the computer programs and / or modules, and the processor implements various functions of the terminal device by running or executing the computer programs and / or modules stored in the memory and calling the data stored in the memory.

[0125] If the module / unit integrated in the terminal device is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.

[0126] Therefore, the present invention adopts the above-mentioned knee exoskeleton multi-action torque assisted control method based on activity classification and phase detection, which can accurately classify the four activity types of standing, walking, climbing stairs and descending stairs, and accurately estimate the gait phase percentage under non-rhythmic gait conditions, and generate the corresponding auxiliary torque curve based on the current gait category and gait phase.

[0127] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A multi-action torque-assisted control method for a knee exoskeleton based on activity classification and phase detection, characterized in that: The following steps are involved: The kinematic data of both sides of the thigh, both sides of the calf and the chest are collected as the input vector X of the artificial neural network and the human body dynamics model; Artificial neural network uses sliding time window method to extract features to achieve action classification; The actual ground gait phase percentage used for model training is obtained by installing foot switches on the soles of the feet; Establish an activity classification neural network that takes the kinematic data of both sides of the thigh as input and outputs the current activity type label; A gait phase detection neural network is established, with the kinematic data of both sides of the thigh as input. The gait phase detection network corresponding to the activity type is selected according to the activity type label output by the activity classification neural network, and the output is the current real-time gait phase. Evaluate activity classification neural networks and gait phase detection neural networks; A human body dynamics model is created. The human body dynamics model generates the real joint biological torque at the current moment according to the current gait phase, and the controller provides an auxiliary torque proportional to the real torque. According to different gait phase modes, the gait phase is divided into single support phase, double support phase and swing phase, and corresponds to three different human body dynamics models. The kinematic data of both sides of the thigh, both sides of the calf and the chest are used as the input of the model to obtain the real biological joint torque at the current moment.

2. The method for multi-action torque-assisted control of a knee joint exoskeleton based on activity classification and phase detection according to claim 1 is characterized in that: Kinematic data include joint angles, angular velocities, and accelerations.

3. The method for multi-action torque-assisted control of a knee joint exoskeleton based on activity classification and phase detection according to claim 1 is characterized in that: Get true ground gait phase percentage by installing foot switches on the sole of the foot, including Each time the foot switch detected ground contact, the gait phase percentage was reset to 0%, the gait phase at the previous time point was set to 100%, and the data points between 0% and 100% were uniformly incremented to obtain the true gait phase percentage at each time point; In the case where the actual duration is shorter than the estimated time, the gait phase percentage is reset to 0 before reaching 100%; if the actual duration is longer than the estimated time, the gait phase percentage remains at 100% until the next ground contact is detected.

4. The method for multi-action torque-assisted control of a knee joint exoskeleton based on activity classification and phase detection according to claim 1 is characterized in that: The activity classification neural network uses a 4×1 vector as output and outputs the result of the classifier as a 4×1 binary vector, corresponding to the four activity types of standing, walking, climbing stairs, and descending stairs. The value of the corresponding activity type is set to 1, and the other activity types are 0.

5. The method for multi-action torque-assisted control of a knee joint exoskeleton based on activity classification and phase detection according to claim 1 is characterized in that: The gait phase detection neural network includes three independent networks, corresponding to three specific typical activity types: walking, climbing stairs and descending stairs.

6. The method for multi-action torque-assisted control of a knee joint exoskeleton based on activity classification and phase detection according to claim 1 is characterized in that: Current real-time gait phase, including: The timestamp between two foot strikes is assigned a gait phase percentage in the range of 0-100%. Since the gait phase percentage jumps from 100% to 0% when the foot strikes the ground, the output one-dimensional result is mapped to a two-dimensional coordinate system by mapping the gait phase percentage to a two-dimensional coordinate (x, y); the two neurons in the output layer are used to output the two-dimensional mapping result of the gait phase percentage; The current gait phase percentage is φ, and the mapping formula is: x=cos(2πφ), y=sin(2πφ)(1); Among them, x represents the abscissa of the two-dimensional coordinate system, and y represents the ordinate of the two-dimensional coordinate system; The calculated two-dimensional mapping (x, y) is calculated by the following equation to obtain the estimated gait phase percentage φ at the current moment current :

7. The method for multi-action torque-assisted control of a knee joint exoskeleton based on activity classification and phase detection according to claim 1, characterized in that: Evaluation of activity classification neural network, gait phase detection neural network, including (1) Evaluation of activity classification neural networks Use the timestamp to describe the precision r a and the detection delay Δt of the activity transition t To evaluate the activity classification network, the formula is as follows: r a =T c / T(3); Among them, T c Indicates the number of correctly classified timestamps, T is the total number of timestamps, N t represents the total number of activity transitions, t t,n and Represents the detected value and true value of the nth activity transition time; (2) Evaluation of gait phase detection neural network The evaluation index of gait phase percentage detection is the relative root mean square error rRMSE of the gait phase percentage, and its specific implementation formula is as follows: in, is the gait phase percentage calculated by (2), φ t is the true value of the gait phase percentage at time t, Represents the mean of the gait phase percentages.

8. The method for multi-action torque-assisted control of a knee joint exoskeleton based on activity classification and phase detection according to claim 1 is characterized in that: The gait phase is divided into single support phase, double support phase and swing phase, and the dynamic equations corresponding to three different human dynamic models are as follows: Single-support human body model: The human body only needs one foot to touch the ground. By setting the weight and inertia parameters of the upper body, thigh and calf, the corresponding inertia matrix Dsp(θ)∈1×5, centrifugal force and Coriolis matrix in the single-support model are derived. And the gravity matrix Gsp(θ)∈1×5, the knee joint torque is derived by the following formula (6): Double support human body model: The human body needs to touch the ground with both feet. The knee joint torque is derived from formula (7). The corresponding inertia matrix in the double support model is Ddp(θ)∈1×5, the centrifugal force and Coriolis matrix The gravity matrix is ​​Gdp(θ)∈1×5. The knee joint moment is derived by the following formula (7): Swinging human body model: The legs are in the swing phase, and the knee joint torque is derived from formula (8). The corresponding inertia matrix in the swinging model is Ds(θ)∈1×5, the centrifugal force and the Coriolis matrix The gravity matrix is ​​Gs(θ)∈1×5, and the knee joint torque is derived by the following formula (8):

9. A multi-action torque-assisted control system for a knee joint exoskeleton based on activity classification and phase detection, characterized in that: It includes artificial neural network and human body dynamics model, and the artificial neural network is divided into the first stage neural network and the second stage neural network; The first-stage neural network consists of an activity classification neural network, and the second-stage neural network consists of three independent gait phase detection neural networks. The activity classification neural network and the gait phase detection neural network are composed of four layers, namely, an input layer, two hidden layers, and an output layer. The first layer is the input layer, which includes 84 neurons; the second hidden layer is the encoder layer, which includes 20 neurons; the third hidden layer is the decoder layer, which includes 40 neurons. The fourth output layer of the activity classification neural network contains 4 neurons, which represent the four activity types of standing, walking, climbing stairs and descending stairs respectively; the first three layers of the gait phase detection neural network have the same structure as the activity classification neural network, and the 2 neurons in the fourth output layer are used to output the gait phase at the current moment.

10. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

Images