Real-time abnormal gait detection system, its training method, and rehabilitation exoskeleton device

Abstract

The present invention belongs to the technical field related to rehabilitation medical assistance, and discloses an abnormal gait real-time detection system, its training method, and a rehabilitation exoskeleton device. The system includes a data acquisition module for constructing a multi-dimensional time series based on the acquired data; a pathological data acquisition module for obtaining the pathological characteristics of a patient; a hybrid normalization processing module for guiding the normalization processing of the sequence according to the pathological characteristics; a low-pass filter bank design module for guiding the learning of the cut-off frequencies of corresponding low-pass filters according to the pathological characteristics; an adaptive filtering module for filtering by using the filters to obtain an enhanced sequence; and an identification module for passing the sequence through a deep learning network to obtain a gait recognition result corresponding to the lower limb. The present invention uses the pathological characteristics of the patient to guide the feature processing and realizes learnable adaptive filtering, which can effectively improve the gait recognition accuracy and enable the rehabilitation exoskeleton device to provide precise assistance for the patient.

Description

Technical Field

[0001] The present invention belongs to the technical field related to rehabilitation medical assistance, and more specifically, relates to an abnormal gait real-time detection system, its training method, and a rehabilitation exoskeleton device. Background Art

[0002] Human gait is a periodic cycle, usually defined as a complete gait cycle from the touchdown of one heel to the next touchdown of the same heel. However, for patients with lower limb motor dysfunction (such as foot drop or hemiplegic patients), due to abnormal neuromuscular control, the gait cycle often exhibits significant asymmetry and spatio-temporal feature disorders. Such abnormal gait is usually accompanied by problems such as blurred cycle boundaries, lag or advance of phase events, which greatly affects the accuracy of gait analysis and restricts the application of rehabilitation exoskeleton devices in dynamic adaptive assistance.

[0003] At present, significant progress has been made in the research on gait phase and event detection. For example, the gait phase estimation algorithm based on long short-term memory network proposed by Mustafa Sarshar et al. realizes high-precision gait phase estimation through triaxial inertial sensor data at the ankle; the time-delay neural network method developed by Shushtari et al. demonstrates excellent robustness and accuracy; the gait phase estimation method based on bidirectional LSTM proposed by Lee et al. is applicable to the control of robotic prosthetics at different walking speeds; Kang et al. proposed a gait phase estimation method based on convolutional neural network and successfully applied it to the real-time control of robotic hip exoskeletons in multi-mode movements. However, the above research has only achieved significant results in healthy populations, but there are difficulties in identifying dynamic gait abnormalities (such as dragging movements) caused by dysfunctions such as foot drop and insufficient hip internal rotation. Summary of the Invention

[0004] In view of the above-mentioned defects or improvement requirements of the prior art, the present invention provides an abnormal gait real-time detection system, its training method, and a rehabilitation exoskeleton device, aiming to accurately achieve gait detection of abnormal gait to provide precise assistance for patients with motor disorders.

[0005] To achieve the above object, according to the first aspect of the present invention, an abnormal gait real-time detection system is provided, which includes:

[0006] A data acquisition and obtaining module for obtaining gait-related data of the same lower limb of a patient to construct multi-dimensional features, and multi-dimensional features at multiple consecutive moments constitute a multi-dimensional time series X e ;

[0007] A pathological data acquisition module for obtaining patient pathological features Y, where the pathological features are a score set of different evaluation indicators of gait;

[0008] A hybrid normalization processing module is used to calculate the ratio of the total score of feature Y to the highest total score of each evaluation index as the pathological weight λ, and perform hybrid normalization processing on each dimension of the sequence X e in the sequence X u to obtain the sequence X u , and the hybrid normalization processing includes making the feature obey the standard normal distribution to obtain the first processing result, making the feature amplitude ratio normalized to obtain the second processing result, and weighted summing the two processing results. The weight of the first processing result is positively correlated with λ, and the weight of the second processing result is negatively correlated with λ;

[0009] A low-pass filter bank design module is used to map feature Y through M fully connected layer networks respectively to obtain the dynamic frequency adjustment amounts of M groups of low-pass filters, and fuse the dynamic frequency adjustment amounts of each low-pass filter in the filter bank with its base frequency to obtain the cut-off frequencies of each low-pass filter in the corresponding filter bank, where M>1 is the feature dimension;

[0010] An adaptive filtering module is used to perform convolution processing on the sequence X u using M groups of low-pass filters. The M groups of low-pass filters correspond one by one to M dimensions. Each group of low-pass filters processes the feature sequence of the corresponding dimension. The feature sequences of each dimension are respectively subjected to convolution calculations through N low-pass filters in the corresponding low-pass filter bank to obtain N convolution sequences of this dimension. The N convolution sequences of the same dimension are first passed through a channel attention network and then fused to obtain the enhanced sequence of this dimension, where N≥1;

[0011] An identification module is used to make the multi-dimensional time series X z composed of the enhanced sequences based on M dimensions pass through a deep learning network to obtain the gait recognition result corresponding to the lower limb.

[0012] Optionally, the gait-related data includes knee joint angle, ankle joint angle, and triaxial angular velocity and triaxial acceleration collected at specific positions, and the specific positions include the front side of the thigh, the front side of the calf, and the instep.

[0013] Optionally, different evaluation indexes of gait include starting, lifting height, step length, gait symmetry, step continuity, walking path, trunk stability, and step width.

[0014] Optionally, a gating mechanism is used to fuse the dynamic frequency adjustment amount of each low-pass filter in the filter bank with its base frequency to obtain the cut-off frequency of each low-pass filter in the corresponding filter bank. The fusion formula is:

[0015] f n c = sigmoid(f n a +α n ⋅ f n ')⋅(fmax -f min ) + f min ;

[0016] where f n c is the cut-off frequency of the nth low-pass filter of any low-pass filter bank, f n a is the base frequency of the nth low-pass filter, f n ' is the dynamic frequency adjustment amount of the nth low-pass filter, α n is a trainable scaling factor, f max , f min are respectively the set maximum and minimum frequencies, and sigmoid() is the activation function.

[0017] Optionally, the system further includes:

[0018] An adaptive adjustment module, configured to map the feature Y to the feature h2 through a fully connected layer network, linearly map the feature h2 through two linear mapping networks respectively to obtain an M-dimensional feature amplitude adjustment amount ρ and an M-dimensional baseline offset correction amount ε, and use the adjustment amount ρ to weight the sequence X f and then fuse it with the sequence X f and the baseline offset correction amount ε to obtain the sequence X Z ; wherein, the sequence X f is directly composed of M-dimensional enhanced sequences, and the ith component in the adjustment amount ρ serves as the weighting weight of the ith-dimensional enhanced sequence in X f , and each eigenvalue of the ith-dimensional enhanced sequence in X f is summed with the ith component in the correction amount ε during fusion.

[0019] Optionally, the gait recognition includes single-task recognition or multi-task recognition;

[0020] The single-task recognition includes: classification of the stance phase and the swing phase, or recognition of continuous gait phase parameters;

[0021] The multi-task recognition includes: classification of the stance phase and the swing phase, and recognition of continuous gait phase parameters.

[0022] Optionally, the gait recognition result includes recognition of continuous gait phase parameters, and the deep learning network outputs the gait phase parameters in the form of a two-dimensional polar coordinate vector.

[0023] According to the second aspect of the present invention, there is provided a training method for the abnormal gait real-time detection system described in the first aspect, which includes:

[0024] Obtain a training data set, where each set of training data includes the pathological characteristics of a patient and the gait-related data of that patient;

[0025] Add gait labels to the gait-related data in each set of training data;

[0026] Input the training data into the system;

[0027] Obtain gait-related data of the same lower limb of the patient through the data acquisition module in the system to construct multi-dimensional features, and the multi-dimensional features at multiple consecutive moments constitute a multi-dimensional time series X e ;

[0028] Obtain the patient's pathological characteristics Y through the pathological data acquisition module in the system, and the pathological characteristics are the score sets of different evaluation indicators of gait;

[0029] Calculate the ratio of the total score of feature Y to the highest total score of each evaluation indicator as the pathological weight λ through the hybrid normalization processing module in the system, and perform hybrid normalization processing on each dimension feature in sequence X e to obtain sequence X u , and the hybrid normalization processing includes making the feature obey the standard normal distribution to obtain the first processing result, making the feature amplitude ratio normalized to obtain the second processing result, and weighted summing the two processing results. The weight of the first processing result is positively correlated with λ, and the weight of the second processing result is negatively correlated with λ;

[0030] Make feature Y pass through M fully connected layer networks through the low-pass filter bank design module in the system to map to the dynamic frequency adjustment amounts of M groups of low-pass filters, and fuse the dynamic frequency adjustment amounts of each low-pass filter in the filter bank with its basic frequency to obtain the cut-off frequencies of each low-pass filter in the corresponding filter bank, where M>1 is the feature dimension;

[0031] Perform convolution processing on sequence X through the adaptive filtering module in the system using M groups of low-pass filters u , where M groups of low-pass filters correspond one-to-one with M dimensions, and each group of low-pass filters processes the feature sequence of the corresponding dimension. The feature sequences of each dimension are respectively convolved by N low-pass filters in the corresponding low-pass filter bank to obtain N convolution sequences of that dimension. The N convolution sequences of the same dimension are first passed through a channel attention network and then fused to obtain the enhanced sequence of that dimension, where N≥1;

[0032] Make the multi-dimensional time series X composed of the enhanced sequences based on M dimensions pass through a deep learning network through the recognition module in the system z to obtain the gait recognition result of the corresponding lower limb;

[0033] Calculate the loss between the gait recognition result and the gait label to train the system until the loss between the gait recognition result output by the system and the gait label converges.

[0034] Optionally, the gait recognition result includes the classification results of the stance phase and the swing phase; the gait label includes the classification labels of the stance phase and the swing phase.

[0035] The process of adding the classification labels of the stance phase and the swing phase to the gait-related data of a certain patient includes:

[0036] Obtain the sagittal plane angular velocity sequence collected from this patient and calculate its mean value ξ and maximum value ω max ;

[0037] Generate an adaptive coefficient β(λ)=β according to the pathological weight λ of this patient min +(β max - β min )⋅λ; β min and β max are set values respectively; the pathological weight λ is the ratio of the total score of the pathological characteristics of this patient to the highest total score of each evaluation index.

[0038] Calculate the angular velocity detection threshold ω(λ)=(ω max ⋅β(λ)+ ξ) / 2;

[0039] Combine the sagittal plane angular velocity curve and the angular velocity detection threshold ω(λ) to identify the heel strike moment HS and the toe off moment TO. Find the maximum value point exceeding the angular velocity detection threshold ω(λ) in the curve as the threshold point. Take the first minimum value point before the threshold point in the curve as the moment TO, and the first minimum value point after the threshold point as the moment HS. If the current gait-related data is in the time interval from the moment TO to the moment HS, it is marked as the stance phase. If the current gait-related data is in the time interval from the moment HS to the moment TO, it is marked as the swing phase.

[0040] According to the third aspect of the present invention, a rehabilitation exoskeleton device is provided, which includes a multi-node wearable data acquisition system, a driving device and a controller. The controller includes the abnormal gait real-time detection system as described in the first aspect;

[0041] The multi-node wearable data acquisition system is used to collect the gait-related data of the patient and transmit it to the controller; the controller is used to input the gait-related data and the pathological data of this patient into the abnormal gait real-time detection system, and control the driving device to assist the corresponding lower limb movement after obtaining the current gait recognition result.

[0042] Generally speaking, compared with the prior art by the above technical solution conceived by the present invention, the present invention mainly has the following beneficial effects.

[0043] 1. In the present invention, the patient's pathological features are introduced to perform mixed normalization processing on the features of the multi-dimensional time series. When the total score of the patient's pathological features is higher, it indicates that the gait regularity of the patient is stronger. At this time, it is more inclined to make the features follow the standard normal distribution to eliminate the distribution shift; when the total score of the patient's pathological features is lower, it indicates that the gait regularity of the patient is weaker. At this time, it is more inclined to perform amplitude ratio normalization on the features to retain the drastic change characteristics of the original amplitude and avoid the Z-score from compressing the information volume of outliers. Therefore, performing the above mixed normalization processing on the collected data helps to improve the gait recognition accuracy and enables the rehabilitation exoskeleton device to provide precise assistance to patients.

[0044] 2. In the present invention, the cut-off frequency of the low-pass filter is a learnable parameter, so that adaptive filtering of the features can be achieved. Through adaptive filtering, the filtering range can be dynamically adjusted according to the patient's individual pathological features, filtering out noise interference while retaining the effective gait information, and more accurately distinguishing the real motion features from the abnormal noise, thereby significantly improving the accuracy of gait phase detection.

[0045] 3. In the present invention, the patient's pathological features are introduced to guide the design of the low-pass filter, and a learnable mapping relationship between the cut-off frequency and the patient's pathological features is established, so as to achieve the adaptation of the spectral characteristics based on the individual pathological state, which helps to improve the gait recognition accuracy and enables the rehabilitation exoskeleton device to provide precise assistance to patients.

[0046] 4. In one embodiment, the dynamic frequency adjustment amount of each low-pass filter in the filter bank is fused with its base frequency through a gating mechanism to obtain the cut-off frequency of each low-pass filter in the corresponding filter bank. In this way, the adaptive regulation of the cut-off frequency can be achieved according to the patient's pathological features. This mechanism not only retains the physical meaning of the base frequency but also generates personalized frequency band selection in combination with the degree of the patient's gait disorder, effectively suppressing the high-frequency noise and motion interference in signal acquisition, while retaining the time-domain features of key gait phase events and improving the signal-to-noise ratio of abnormal gait signals.

[0047] 5. In one embodiment, an adaptive adjustment module is set up to adaptively adjust the enhanced features by introducing pathological features. In this way, the signal features can be adaptively optimized according to the degree of the individual's gait disorder, so as to more accurately capture the subtle changes of abnormal gaits, enhancing the model's ability to distinguish complex gait patterns.

[0048] 6. In one embodiment, the deep learning network outputs gait phase parameters in the form of two-dimensional polar coordinate vectors. This mapping method converts the linearly increasing phase parameter θ into a continuously varying periodic signal, eliminating the jump problem at the HS event. Thus, it is beneficial for system training, helps improve the accuracy of gait recognition, and enables the rehabilitation exoskeleton device to provide precise assistance to patients.

[0049] 7. In one embodiment, introducing the patient's pathological characteristics to guide the determination of the angular velocity detection threshold can adaptively adjust the threshold according to the patient's pathological condition. Thereby, the moments of heel strike and toe off can be marked more precisely, the stance phase and swing phase can be accurately divided, and then a more precise real-time abnormal gait detection system can be trained, enabling the rehabilitation exoskeleton device to provide precise assistance to patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a block diagram of the real-time abnormal gait detection system in one embodiment of the present invention.

[0051] Figure 2 is a block diagram of the real-time abnormal gait detection system in another embodiment of the present invention.

[0052] Figure 3 is a schematic diagram of the wearable data acquisition system for gait estimation in one embodiment of the present invention.

[0053] Figure 4 is a schematic diagram of the structure of the adaptive filtering module in one embodiment of the present invention.

[0054] Figure 5 is a comparison diagram of gait labels in the form of continuous gait phase parameters and two-dimensional polar coordinate vectors in one embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0055] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0056] Embodiment 1

[0057] The present invention provides a real-time abnormal gait detection system. As Figure 1 shown is a block diagram of the real-time abnormal gait detection system in one embodiment of the present invention, which includes a collected data acquisition module, a pathological data acquisition module, a hybrid normalization processing module, a low-pass filter bank design module, an adaptive filtering module, and an identification module. Each of the modules will be described in detail below.

[0058] The data acquisition module is used to acquire gait-related data of the same lower limb of a patient to construct multi-dimensional features, and the multi-dimensional features at multiple consecutive moments constitute a multi-dimensional time series X e 。

[0059] The present invention adopts a single-leg detection method. If the left leg needs to be assisted, the left leg data is collected, and the left leg gait is detected based on the left leg data. If the right leg needs to be assisted, the right leg data is collected, and the right leg gait is detected based on the right leg data. If both legs need to be assisted, the data of both legs is collected simultaneously, the left leg gait is detected based on the left leg data, and the right leg gait is detected based on the right leg data.

[0060] Specifically, in order to improve the detection accuracy, information at multiple positions is collected. The gait-related data may include, for example, knee joint angle, ankle joint angle, and three-axis angular velocity and three-axis acceleration collected at specific positions. The specific positions include the front side of the thigh, the front side of the calf, and the instep, etc.

[0061] Specifically, before data processing, the signals at the front side of the thigh, the front side of the calf, and the instep of the same lower limb are collected first. In one embodiment, in order to obtain high-precision gait phase detection data, a multi-node wearable data acquisition system is built, such as Figure 3 The figure shows a schematic diagram of a wearable data acquisition system for gait estimation in an embodiment of the present invention. Taking assisting both legs as an example, the system uses six micro inertial measurement units (IMUs) as core sensors, which are respectively installed at the front side of the thigh, the front side of the calf, and the instep of the subject's two legs to achieve a comprehensive capture of the lower limb movement information. Each IMU records its quaternion, three-axis angular velocity, and three-axis acceleration in real time at the same sampling frequency (such as 100Hz), and synchronously communicates with the controller through a wireless transmission module to transmit the collected data to the controller.

[0062] Among them, based on the quaternion data collected by the 3 IMUs on the same lower limb, through the attitude relationship between adjacent body segments, the angles of the knee joint and ankle joint of this lower limb can be accurately calculated by using the quaternion difference method.

[0063] At this time, based on the gait-related data of the same lower limb, a multi-dimensional feature X(t) containing M dimensions at time t for this lower limb can be obtained. The multi-dimensional features X(t) at multiple consecutive moments constitute a multi-dimensional time series, X e ∈R T╳M , T is the number of moments, that is, the sequence length. The multi-dimensional feature at each moment may include the three-axis angular velocity (9 dimensions) and three-axis acceleration (9 dimensions) collected by the 3 parts of the IMU at this moment, as well as the knee joint angle (1 dimension) and ankle joint angle (1 dimension) obtained by calculation.

[0064] Further, after the signal is collected, preprocessing can be performed first, such as performing simple denoising on the IMU data, and then performing Z-Score normalization on the acceleration and angular velocity.

[0065] The pathological data acquisition module is used to obtain the patient's pathological feature Y, and the pathological feature is a score set of different evaluation indexes of gait.

[0066] Since the degrees of motor impairment of different patients are different, that is, there are differences between individuals. In the present invention, the patient's pathological feature Y is obtained, and the pathological feature Y can reflect the degrees of motor impairment of different individuals. Introducing the pathological feature Y can improve the accuracy of gait recognition.

[0067] Specifically, the patient's pathological feature Y can be constructed based on the patient's Tinetti gait score form. Different evaluation indexes of gait include starting, lifting height of the foot, step length, gait symmetry, step continuity, walking path, trunk stability, and step width.

[0068] Experienced doctors evaluate the patient's gait ability according to the Tinetti gait score form. The doctors test the patient through 8 gait test items including starting, lifting height of the foot, step length, gait symmetry, step continuity, walking path, trunk stability, and step width, and score each item. The score value can be set to (0, 1, 2), where a higher score indicates better gait ability of the patient. Thus, 8 parameters of the gait ability of each patient are obtained as the patient's pathological feature Y = [y1, y2, …, y8] ∈ R 8 , where, y k is the score of the kth index.

[0069] The hybrid normalization processing module is used to calculate the ratio of the total score value of the feature Y to the highest total score value of the pathological feature as the pathological weight λ, and perform hybrid normalization processing on each dimension feature in the sequence X e to obtain the sequence X u . The hybrid normalization processing includes making the feature obey the standard normal distribution to obtain the first processing result and making the feature amplitude ratio normalized to obtain the second processing result and weighted summing the two processing results. The weight of the first processing result is positively correlated with λ, and the weight of the second processing result is negatively correlated with λ.

[0070] Specifically, λ = S / S max , , S max is the highest total score value of the pathological feature, that is, the summation result after each index takes the highest score.

[0071] Specifically, the hybrid normalization processing can be written in the following function form:

[0072] xnorm = λ ⋅ ((x - μ) / σ) + (1 - λ) ⋅ (x / (c ⋅ max));

[0073] Where x is the sequence X e Among them, μ, σ, and max, which are one of the data in any dimension, are the mean, variance, and maximum value of all data in the dimension to which x belongs, respectively, and c is a preset scale factor.

[0074] In fact, the first term in the above function is a transformation function that makes the features follow a standard normal distribution, and the second term is a transformation function that normalizes the feature amplitude ratio. Based on this hybrid normalization process, when the total score S of the patient's pathological feature Y is higher, it indicates that the patient's gait regularity is stronger. At this time, more emphasis is placed on making the features follow a standard normal distribution to eliminate distribution deviation. When the total score S of the patient's pathological feature Y is lower, it indicates that the patient's gait regularity is weaker. At this time, more emphasis is placed on normalizing the feature amplitude ratio to retain the sharp change characteristics of the original amplitude and avoid the Z-score compressing the information of outliers. Therefore, performing the above hybrid normalization process on the collected data helps to improve the accuracy of gait recognition.

[0075] The low-pass filter bank design module is used to map the feature Y through M fully connected layer networks respectively to obtain the dynamic frequency adjustment amounts of M groups of low-pass filters, and fuse the dynamic frequency adjustment amounts of each low-pass filter in the filter bank with its base frequency to obtain the cut-off frequencies of each low-pass filter in the corresponding filter bank, where M is the feature dimension.

[0076] Specifically, the low-pass filter bank design module includes M parallel fully connected layer networks. Each fully connected layer network corresponds to the design of a group of low-pass filters. When the feature Y is input into the fully connected layer network for feature mapping, the dynamic frequency adjustment amounts of each low-pass filter in the corresponding low-pass filter combination can be obtained.

[0077] In one embodiment, each fully connected layer network includes two layers of fully connected layers. In each fully connected layer network, the feature Y ∈ R 8 is first mapped to a high-dimensional hidden space through a fully connected layer to obtain an intermediate representation h1 ∈ R 32 , and then the intermediate representation h1 is mapped through a fully connected layer to obtain the dynamic frequency adjustment amount f' = [f1', f2'... f N '], where N is the number of low-pass filters in the corresponding low-pass filter bank, and f n ' is the dynamic frequency adjustment amount of the nth low-pass filter in the corresponding filter bank.

[0078] Specifically, the process of obtaining the intermediate representation h1 can be written as the following expression:

[0079] h1 = ReLU(W1Y + b1);

[0080] where \(W_1\in\mathbb{R}\) 32╳8 is a trainable weight matrix, \(b_1\in\mathbb{R}\) 32 is a trainable bias term, and ReLU() represents the activation function.

[0081] Specifically, the process of obtaining the dynamic frequency adjustment amount \(f'\) can be written as the following expression:

[0082] \(f' = W_2h_1 + b_2\);

[0083] where \(W_2\in\mathbb{R}\) N╳32 is a trainable weight matrix, \(b_2\in\mathbb{R}\) N is a trainable bias term.

[0084] After obtaining the dynamic frequency adjustment amount of each low - pass filter, the dynamic frequency adjustment amount of the low - pass filter is fused with its set base frequency to obtain the cut - off frequency of the corresponding low - pass filter.

[0085] In one embodiment, the dynamic frequency adjustment amount of the low - pass filter can be directly added to its set base frequency to obtain the cut - off frequency of the corresponding low - pass filter.

[0086] In another embodiment, the dynamic frequency adjustment amount of the low - pass filter can also be fused with its set base frequency through a gating mechanism. Taking the \(n\)th low - pass filter of any low - pass filter bank as an example, its gating mechanism is used to fuse the dynamic frequency adjustment amount with its base frequency to obtain the cut - off frequency of this low - pass filter. The fusion process can be written as the following expression:

[0087] \(f\) n c \(=\text{sigmoid}(f\) n a +\alpha\) n \(\cdot f\) n ')\cdot(f\) max -f\) min ) + f\) min ;

[0088] where \(f\) n c is the cut - off frequency of the \(n\)th low - pass filter of any low - pass filter bank, \(f\) n a is the base frequency of the \(n\)th low - pass filter, \(f\) n ' is the dynamic frequency adjustment amount of the \(n\)th low - pass filter, \(\alpha\) n is a trainable scaling factor used to dynamically adjust the sensitivity so that the cut - off frequency can be dynamically adapted to the pathological features, \(f\) max and \(f\) minare the set maximum and minimum frequencies respectively, that is, the frequency band range is preset. sigmoid() is an activation function that maps unbounded inputs to the interval (0, 1), thus scaling the cut-off frequency to the interval (f max , f min ).

[0089] In this embodiment, frequencies are fused through a gating mechanism. In this way, the adaptive regulation of the cut-off frequency can be achieved according to the patient's pathological characteristics. This mechanism not only retains the physical meaning of the basic frequency but also generates personalized frequency band selection in combination with the degree of the patient's gait disorder, effectively suppressing high-frequency noise and motion interference in signal acquisition, while retaining the time-domain characteristics of key gait phase events and improving the signal-to-noise ratio of abnormal gait signals.

[0090] It should be noted that the basic frequency of the low-pass filter can be set according to prior knowledge of the gait cycle and event characteristics, and can be set to be the same or different specifically.

[0091] In one embodiment, the basic frequencies of the N low-pass filters in each low-pass filter bank can be set to be uniformly distributed in the interval (f max , f min ). In this way, a multi-scale filtering network covering the target frequency domain range can be constructed in the initial design stage, and the dynamic characteristics of different frequency components in the gait signal can be captured through the differential distribution of the basic frequencies.

[0092] It can be understood that once the cut-off frequency of the low-pass filter is determined, the time-domain impulse response function of the low-pass filter can be determined, thereby realizing the design of the low-pass filter.

[0093] Taking the nth low-pass filter of any low-pass filter bank as an example, the expression of the time-domain impulse response of the low-pass filter is as follows:

[0094] h(t)=2f n c ·sinc(2πf n c t);

[0095] In the formula, f n c is the nth low-pass filter of any low-pass filter bank, t is time, and h(t) is the response of this low-pass filter.

[0096] In one embodiment, each group of low-pass filter banks may only contain one low-pass filter.

[0097] In another embodiment, each group of low-pass filter banks may contain multiple low-pass filters, thereby realizing multi-scale adaptive filtering.

[0098] In one embodiment, the above original time-domain impulse response can be directly used as the actual response of the low-pass filter.

[0099] In another embodiment, further processing can also be performed on the original time-domain impulse response. To suppress spectral leakage, first, a Hamming window function Q(t) = 0.54 - 0.46 cos(2πt / T') with a length of T' is used to window the original impulse response h(t), and the windowed finite impulse response h'(t) = h(t) ⋅ Q(t) is obtained. Subsequently, to ensure the symmetry of the time-domain response of the low-pass filter, a symmetric impulse response h s '(t) is generated by mirror extension, and normalization is performed to obtain the normalized response h norm '(t) = h s '(t) / ∑ t h s '(t), which is used as the actual response of the low-pass filter.

[0100] In this embodiment, by windowing, symmetrizing, and normalizing the original time-domain impulse response, the transition band fluctuation of the frequency response curve can be smoothed, the time-domain shift of gait key features can be avoided, the balance of multi-channel signal energy distribution can be improved, and the stability and consistency of feature extraction in a motion interference scenario can be enhanced.

[0101] The adaptive filtering module is used to perform convolution processing on the sequence X u using M groups of low-pass filters. The M groups of low-pass filters correspond one-to-one with M dimensions. Each group of low-pass filters processes the feature sequence of the corresponding dimension. The feature sequences of each dimension are respectively convolved by N low-pass filters in the corresponding low-pass filter group to obtain N convolution sequences of this dimension. The N convolution sequences of the same dimension are first passed through a channel attention network and then fused to obtain the enhanced sequence of this dimension.

[0102] As Figure 4 shown in the structural schematic diagram of the adaptive filtering module in an embodiment of the present invention, the adaptive filtering module processes the feature sequences of each dimension in the multi-dimensional time series X u respectively to achieve feature enhancement.

[0103] Specifically, the adaptive filtering module has M parallel adaptive filtering sub-modules. The M groups of low-pass filters are distributed one-to-one in the M sub-modules. The M sub-modules correspond one-to-one with the M dimensions in the multi-dimensional time series X u , that is, each sub-module is used to process the feature sequence of the corresponding dimension in the multi-dimensional time series X u .

[0104] The process of feature enhancement of the feature sequence of each dimension through its corresponding sub-module includes:

[0105] The feature sequence is first convolved through N low-pass filters in the sub-module to obtain N convolutional sequences, and the N convolutional sequences are X' ∈ R T╳N ;

[0106] X' is enhanced in features through a channel attention network to obtain N intermediate sequences;

[0107] The eigenvalue at the same moment in the N intermediate sequences is superimposed to fuse the N intermediate sequences into one sequence to obtain an enhanced sequence.

[0108] Among them, the specific working process of the channel attention network is as follows:

[0109] First, global average pooling is performed on the sequence X' to compress the spatio-temporal dimension to obtain a channel description vector z = GAP(X'), z ∈ R N , where GAP() is the global average pooling function;

[0110] Subsequently, the pooling result z is processed using two fully connected networks, and the attention weight distribution g = σ(W4 ⋅ GELU(W3z)) for N channels (i.e., N convolutional sequences) is calculated. W3 and W4 are the weight matrices of the two fully connected networks respectively, and GELU() and σ() are the activation functions of the two fully connected networks;

[0111] Finally, each convolutional sequence in X' is first weighted with the corresponding channel attention weight and then fused with the convolutional sequence to obtain N intermediate sequences.

[0112] After obtaining the N intermediate sequences, feature fusion is performed again to obtain an enhanced sequence in the corresponding dimension.

[0113] Specifically, the calculation formula for the enhanced sequence in each dimension finally obtained can be expressed as:

[0114] ;

[0115] where X f,j is the enhanced sequence in the jth dimension, X i ' is the convolutional sequence of the input i-th channel in the sequence X', and g i is the attention weight corresponding to the i-th channel in the attention weight distribution g.

[0116] Finally, directly combine the enhanced sequences of M dimensions, and then a multi-dimensional enhanced time series X f ∈ R T╳M .

[0117] The recognition module is used to make the multi-dimensional time series X composed of the enhanced sequences of M dimensions zThrough the deep learning network, the gait recognition result corresponding to the lower limbs is obtained.

[0118] Specifically, the choice of the deep learning network is not limited and can be a transformer, RNN, CNN, or LSTM (Long Short-Term Memory).

[0119] In this embodiment, a bidirectional LSTM can be selected. By simultaneously capturing the forward and backward dependencies of the time series, the bidirectional LSTM can more comprehensively model the dynamic characteristics of the gait.

[0120] In one embodiment, the multi-dimensional enhanced time series X obtained by directly combining M-dimensional enhanced sequences can be used f as the input X of the deep learning network z .

[0121] In another embodiment, as shown in Figure 2 , the abnormal gait real-time detection system further includes an adaptive adjustment module;

[0122] The adaptive adjustment module is used to map the feature Y to the feature h2 through a fully connected layer network, and linearly map the feature h2 through two linear mapping networks respectively to obtain the feature amplitude adjustment amount ρ ∈ R M and the baseline offset correction amount ε ∈ R M . The sequence X is weighted by using the adjustment amount ρ f and then fused with the sequence X f and the baseline offset correction amount ε to obtain the sequence X Z ; where the sequence X f ∈R T╳M is composed of M-dimensional enhanced sequences combined, T is the sequence length, and the i-th component in the adjustment amount ρ is used as the weighting weight of the i-th dimensional enhanced sequence in X f . When fusing, each eigenvalue of the i-th dimensional enhanced sequence in X f is summed with the i-th component in the correction amount ε.

[0123] Specifically, the feature Y is mapped to the feature h2 through a fully connected network, and the semantic information of the medical evaluation scale is mapped to the feature h2 to provide an information basis for subsequent parameter separation. The corresponding process can be expressed by the following expression:

[0124] h2 = ReLU(W5Y + b5);

[0125] In the formula, W5 and b5 are the weight matrix and bias of this fully connected layer, and ReLU() is the activation function.

[0126] Let the feature h2 pass through two linear mapping networks for independent linear mapping respectively to obtain the feature amplitude adjustment ρ and the baseline offset correction ε. The calculation formula can be expressed as:

[0127] ρ = W6h2 + b6;

[0128] ε = W7h2 + b7;

[0129] In the formula, W6 and W7 are the weight matrices of the two linear mapping networks respectively, and b6 and b7 are the biases of the two linear mapping networks respectively.

[0130] Use the adjustment ρ to weight the sequence X f and then fuse it with the sequence X f and the baseline offset correction ε to obtain the sequence X Z . The calculation formula can be expressed as:

[0131] X Z = X f + (ρ ⊙ X f + Ω);

[0132] In the formula, ⊙ represents element-wise multiplication, and Ω is a matrix with the same dimension as X f formed by replicating ε T times.

[0133] In this embodiment, pathological features are introduced to adaptively adjust the enhanced features. In this way, the signal features can be adaptively optimized according to the degree of gait disorder of an individual, so as to more accurately capture the subtle changes of abnormal gaits and enhance the resolution ability of the model for complex gait patterns.

[0134] Among them, the gait recognition task can be selected according to requirements.

[0135] In one embodiment, gait recognition includes single-task recognition, and the single-task recognition includes: classification of the stance phase and the swing phase, or recognition of continuous gait phase parameters.

[0136] In another embodiment, gait recognition includes multi-task recognition, and the multi-task recognition includes: classification of the stance phase and the swing phase, and recognition of continuous gait phase parameters. At this time, after feature extraction and sequence modeling are completed, the network performs the classification task of the stance phase and the swing phase (output through the classification head) and the regression task of continuous gait phase parameters (output through the regression head) respectively through two independent fully connected layers. The output of each task is based on the last time step of the sequence to ensure that it can integrate the most representative gait features in the whole sequence.

[0137] Since the current mainstream gait detection methods mainly focus on single-task detection. For example, they only detect continuous gait phases or only determine whether the current gait is in the stance phase or the swing phase. However, in practical applications, multi-information cooperation is required: continuous phases are used to calculate the assistance intensity, while discrete events (stance phase or swing phase) determine the triggering timing. Since existing methods focus on single-task detection, it is difficult to meet the demand for precise assistance of lower-limb exoskeletons in complex rehabilitation scenarios. In this embodiment, dual-task recognition is adopted to meet the demand for precise assistance of lower-limb exoskeletons in complex rehabilitation scenarios.

[0138] In one embodiment, when the gait recognition result includes the recognition of continuous gait phase parameters, the deep learning network outputs the gait phase parameters in the form of two-dimensional polar coordinate vectors.

[0139] Specifically, the gait phase parameter θ is defined as follows: within the gait cycle from one heel strike (HS) event to the next heel strike, the gait phase parameter θ linearly increases from 0 to 1, which is used to represent the relative position of the current moment in the gait cycle. However, as Figure 5 shown, this definition will cause the gait phase parameter θ to instantaneously jump from 1 to 0 at the HS event, forming a discontinuity. This discontinuity will interfere with the training process of the system and affect the convergence and stability of the system.

[0140] To solve this problem, in this embodiment, a polar coordinate-based gait phase representation method is adopted to convert the scalar gait phase parameter θ into the form of two-dimensional polar coordinate vectors (x1, x2), and the expression is:

[0141] (x1, x2) = (cos(2πθ), sin(2πθ));

[0142] After the transformation, the components x1 and x2 of the obtained two-dimensional polar coordinate vector (x1, x2) are both in the range of (-1, 1) and satisfy x1 2 + x2 2 = 1. This mapping method converts the linearly increasing phase parameter θ into a continuously changing periodic signal, eliminating the jump problem at the HS event. In this way, it is beneficial to system training and improves the accuracy of the system.

[0143] Embodiment 2

[0144] The present invention also provides a training method for the abnormal gait real-time detection system described in Embodiment 1, including:

[0145] Obtain a training data set, and each group of training data includes the pathological characteristics of the patient and the gait-related data of this patient;

[0146] Add gait labels to the gait-related data in each set of training data;

[0147] Input the training data into the system;

[0148] Obtain gait-related data of the same lower limb of the patient through the data acquisition module in the system to construct multi-dimensional features, and the multi-dimensional features at multiple consecutive moments constitute the multi-dimensional time series X e ;

[0149] Obtain the patient's pathological features Y through the pathological data acquisition module in the system, and the pathological features are the score sets of different evaluation indicators of the gait;

[0150] Calculate the ratio of the total score of the feature Y to the highest total score of each evaluation indicator as the pathological weight λ through the hybrid normalization processing module in the system, and perform hybrid normalization processing on each dimension feature in the sequence X e to obtain the sequence X u , and the hybrid normalization processing includes making the feature obey the standard normal distribution to obtain the first processing result and making the feature amplitude ratio normalized to obtain the second processing result and weighted summing the two processing results. The weight of the first processing result is positively correlated with λ, and the weight of the second processing result is negatively correlated with λ;

[0151] Make the feature Y pass through M fully connected layer networks through the low-pass filter bank design module in the system to be mapped into the dynamic frequency adjustment amounts of M groups of low-pass filters, and fuse the dynamic frequency adjustment amounts of each low-pass filter in the filter bank with its base frequency to obtain the cut-off frequencies of each low-pass filter in the corresponding filter bank, where M>1 is the feature dimension;

[0152] Perform convolution processing on the sequence X through the adaptive filtering module in the system using M groups of low-pass filters u , where M groups of low-pass filters correspond one-to-one with M dimensions, and each group of low-pass filters processes the feature sequence of the corresponding dimension. The feature sequences of each dimension are respectively convolved by N low-pass filters in the corresponding low-pass filter bank to obtain N convolution sequences of this dimension. The N convolution sequences of the same dimension are first passed through the channel attention network and then fused to obtain the enhanced sequence of this dimension, where N≥1;

[0153] Make the multi-dimensional time series X composed of the enhanced sequences based on M dimensions pass through the recognition module in the system z through the deep learning network to obtain the gait recognition result of the corresponding lower limb;

[0154] Calculate the loss between the gait recognition result and the gait label to train the system until the loss between the gait recognition result output by the system and the gait label converges.

[0155] When gait recognition is the classification of the stance phase and the swing phase, the gait label is the classification label of the stance phase and the swing phase;

[0156] When gait recognition is the recognition of continuous gait phase parameters, the gait label is the continuous gait phase parameter label;

[0157] When gait recognition includes both the classification of the stance phase and the swing phase and the recognition of continuous gait phase parameters, the gait label includes both the classification label of the stance phase and the swing phase and the continuous gait phase parameter label.

[0158] In one embodiment, when the system outputs gait phase parameters in the form of two-dimensional polar coordinate vectors, the continuous gait phase parameter labels used for training are also in the form of two-dimensional polar coordinate vector labels.

[0159] In one embodiment, when gait recognition includes the classification of the stance phase and the swing phase, the gait label includes the classification label of the stance phase and the swing phase. The process of adding the classification label of the stance phase and the swing phase to the gait-related data of a certain patient includes:

[0160] Obtain the sagittal plane angular velocity sequence collected by this patient and calculate its mean value ξ and maximum value ω max ;

[0161] Generate an adaptive coefficient β(λ)=β min +(β max - β min )⋅λ; β min and β max are respectively set positive values, and their optimal values can be determined according to experiments. For example, β min =0.01, β max =0.3; the pathological weight λ is the ratio of the total score of the pathological features of this patient to the highest total score of each evaluation index;

[0162] Calculate the angular velocity detection threshold ω(λ)=(ω max ⋅β(λ)+ ξ) / 2;

[0163] Combine the sagittal plane angular velocity curve and the angular velocity detection threshold ω(λ) to identify the heel strike moment HS and the toe off moment TO. Find the maximum value point exceeding the angular velocity detection threshold ω(λ) in the curve as the threshold point, use the first minimum value point before the threshold point in the curve as the moment TO, and the first minimum value point after the threshold point as the moment HS. If the current gait-related data is in the time interval from the moment TO to the moment HS, it is marked as the stance phase. If the current gait-related data is in the time interval from the moment HS to the moment TO, it is marked as the swing phase.

[0164] Specifically, this method analyzes the angular velocity curve of the shank in the sagittal plane to find its periodic change characteristics. In order to identify whether the current gait-related data corresponds to the swing phase or the stance phase, two types of moments in the curve need to be identified, namely the heel strike moment HS and the toe off moment TO. During the time interval from moment TO to the adjacent moment HS, the lower limb is in the stance phase. During the time interval from moment HS to the adjacent moment TO, the lower limb is in the swing phase.

[0165] In this embodiment, considering that there are significant gait differences among different patients, by introducing the pathological characteristics of the patient to guide the determination of the angular velocity detection threshold, the threshold can be adaptively adjusted according to the patient's pathological condition. Thus, the moments of heel strike and toe off can be marked more accurately, so as to accurately divide the stance phase and the swing phase, and then train a more accurate real-time abnormal gait detection system.

[0166] Embodiment 3

[0167] The present invention also provides a rehabilitation exoskeleton device, which includes a multi-node wearable data acquisition system, a driving device, and a controller. The controller includes the real-time abnormal gait detection system as described in Embodiment 1. Among them, the multi-node wearable data acquisition system is used to collect the gait-related data of the patient and transmit it to the controller. The controller is used to input the gait-related data and the pathological data of the patient into the real-time abnormal gait detection system, and control the driving device to assist the corresponding lower limb movement after obtaining the current gait recognition result.

[0168] Specifically, the driving device can be composed of two motors. Each motor is connected to an independent Bowden cable. The two Bowden cables are respectively connected to the instep and the heel of the wearer. When receiving a control instruction, the motor rotates to drive the Bowden cable to contract or relax, so as to assist the wearer in performing dorsiflexion and plantar flexion, thereby assisting the lower limb movement.

[0169] The technical features of the above-described embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as these combinations of technical features do not conflict, they should be considered to be within the scope described in this specification. It should be noted that "in an embodiment of the present invention", "for example", "similarly" and the like in the present invention are intended to illustrate the present invention rather than to limit the present invention.

[0170] The above-described embodiments only represent several implementation manners of the present invention. Their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the patent application. It should be pointed out that for those of ordinary skill in the art, without departing from the concept of the present invention, several deformations and improvements can be made, and these all belong to the protection scope of the present invention.

Claims

1. A real-time abnormal gait detection system, characterized in that, Including: A data acquisition module, configured to acquire gait-related data of the same lower limb of a patient to construct multi-dimensional features, and the multi-dimensional features at multiple consecutive moments form a multi-dimensional time series X e ; A pathological data acquisition module, configured to acquire a patient's pathological feature Y, where the pathological feature is a score set of different evaluation indexes of gait; A hybrid normalization processing module is used to calculate the ratio of the total score of feature Y to the highest total score of each evaluation index as the pathological weight λ, and perform hybrid normalization processing on each dimension of feature in sequence X e to obtain sequence X u . The hybrid normalization processing includes making the feature obey the standard normal distribution to obtain a first processing result, making the feature amplitude ratio normalized to obtain a second processing result, and weighting and summing the two processing results. The weight of the first processing result is positively correlated with λ, and the weight of the second processing result is negatively correlated with λ; A low-pass filter bank design module, configured to map the feature Y through M fully-connected layer networks respectively to obtain dynamic frequency adjustment amounts of M groups of low-pass filters, and fuse the dynamic frequency adjustment amounts of each low-pass filter in the filter bank with its base frequency to obtain the cut-off frequency of each low-pass filter in the corresponding filter bank, where M>1 is the feature dimension; An adaptive filtering module for convolving the sequence X using M groups of low-pass filters u to perform convolution processing. The M groups of low-pass filters correspond one-to-one with M dimensions. Each group of low-pass filters processes the feature sequence of the corresponding dimension. The feature sequences of each dimension are respectively convolved by N low-pass filters in the corresponding low-pass filter group to obtain N convolution sequences of that dimension. The N convolution sequences of the same dimension are first passed through a channel attention network and then fused to obtain an enhanced sequence of that dimension, where N ≥ 1; An identification module for enabling a multi-dimensional time series X composed of enhanced sequences based on M dimensions z to obtain a gait recognition result corresponding to the lower limb through a deep learning network.

2. The real-time abnormal gait detection system according to claim 1, characterized in that, The gait-related data includes knee joint angle, ankle joint angle, and three-axis angular velocity and three-axis acceleration collected at specific positions, and the specific positions include the front side of the thigh, the front side of the calf, and the instep.

3. The real-time abnormal gait detection system according to claim 1, wherein, The different evaluation indexes of gait include starting, lifting height, step length, gait symmetry, step continuity, walking path, trunk stability, and step width.

4. The real-time abnormal gait detection system according to claim 1, characterized in that, Adopt a gating mechanism to fuse the dynamic frequency adjustment amount of each low-pass filter in the filter bank with its base frequency to obtain the cut-off frequency of each low-pass filter in the corresponding filter bank, and the fusion formula is: f n c = sigmoid(f n a +α n ⋅ f n ')⋅(f max -f min )+f min ; where f n c is the cut-off frequency of the n-th low-pass filter of any low-pass filter bank, f n a is the base frequency of the n-th low-pass filter, f n ' is the dynamic frequency adjustment amount of the n-th low-pass filter, α n is a trainable scaling factor, f max and f min are the set maximum and minimum frequencies respectively, and sigmoid() is the activation function.

5. The real-time abnormal gait detection system according to claim 1, wherein The system further includes: The adaptive adjustment module is used to map feature Y to feature h2 through the fully connected layer network, and to linearly map feature h2 through two linear mapping networks to obtain the M-dimensional feature amplitude adjustment ρ and the M-dimensional baseline offset correction ε. The adjustment ρ is used to adjust the sequence X f After weighting, it is then combined with the sequence X f , and the baseline offset correction ε are fused to obtain the sequence X Z ; where sequence X f is composed of M-dimensional enhanced sequences directly combined, and the i-th component in the adjustment amount ρ is used as X f The weighted weight of the i-th dimension enhanced sequence in , when fused X f Each eigenvalue of the i-th dimension enhanced sequence in is summed with the i-th component of the correction value ε.

6. The real-time abnormal gait detection system according to claim 1, characterized in that, The gait recognition includes single-task recognition or multi-task recognition; The single-task recognition includes: classification of the stance phase and the swing phase, or recognition of continuous gait phase parameters; The multi-task recognition includes: classification of the stance phase and the swing phase, and recognition of continuous gait phase parameters.

7. The real-time abnormal gait detection system according to claim 1, characterized in that, The gait recognition result includes the recognition of continuous gait phase parameters, and the deep learning network outputs the gait phase parameters in the form of a two-dimensional polar coordinate vector.

8. A training method for the abnormal gait real-time detection system according to any one of claims 1 to 7, characterized in that, Including: Obtain a training data set, and each group of training data includes a patient's pathological feature and the patient's gait-related data; Add a gait label to the gait-related data in each group of training data; Input the training data into the system; The gait-related data of the same lower limb of the patient is obtained through the acquisition data obtaining module in the system to construct multi-dimensional features, and the multi-dimensional features at multiple consecutive moments constitute a multi-dimensional time series X e ; Obtain a patient's pathological feature Y through the pathological data acquisition module in the system, where the pathological feature is a score set of different evaluation indexes of gait; Calculate the ratio of the total score of feature Y to the highest total score of each evaluation index as the pathological weight λ through the hybrid normalization processing module in the system, and perform hybrid normalization processing on each dimension of feature in sequence X e to obtain sequence X u . The hybrid normalization processing includes making the feature obey the standard normal distribution to obtain the first processing result, making the feature amplitude ratio normalized to obtain the second processing result, and weighted summing the two processing results. The weight of the first processing result is positively correlated with λ, and the weight of the second processing result is negatively correlated with λ; Map the feature Y through M fully-connected layer networks respectively through the low-pass filter bank design module in the system to obtain dynamic frequency adjustment amounts of M groups of low-pass filters, and fuse the dynamic frequency adjustment amounts of each low-pass filter in the filter bank with its base frequency to obtain the cut-off frequency of each low-pass filter in the corresponding filter bank, where M>1 is the feature dimension; The sequence X is convolved by using M groups of low-pass filters in the adaptive filtering module in the system u The M groups of low-pass filters correspond one-to-one with M dimensions. Each group of low-pass filters processes the feature sequence of the corresponding dimension. The feature sequences of each dimension are respectively convolved by N low-pass filters in the corresponding low-pass filter bank to obtain N convolution sequences of that dimension. The N convolution sequences of the same dimension are first passed through a channel attention network and then fused to obtain an enhanced sequence of that dimension, where N≥1; The multi-dimensional time series X composed of enhanced sequences based on M dimensions is obtained through the recognition module in the system z The gait recognition result corresponding to the lower limb is obtained through a deep learning network; Calculate the loss between the gait recognition result and the gait label to train the system until the loss between the gait recognition result output by the system and the gait label converges.

9. The training method according to claim 8, wherein The gait recognition result includes the classification result of the stance phase and the swing phase; the gait label includes the classification label of the stance phase and the swing phase; The process of adding the classification label of the stance phase and the swing phase to the gait-related data of a certain patient includes: Obtain the sagittal plane angular velocity sequence collected from this patient and calculate its mean value ξ and maximum value ω max ; Generate an adaptive coefficient β(λ) = β according to the pathological weight λ of the patient min +(β max - β min ) ⋅ λ; β min and β max are set values respectively; the pathological weight λ is the ratio of the total score of the pathological features of the patient to the highest total score of each evaluation index; Calculate the angular velocity detection threshold ω(λ) = (ω max ⋅β(λ) + ξ) / 2; Identify the heel strike moment HS and the toe off moment TO by combining the sagittal plane angular velocity curve and the angular velocity detection threshold ω(λ). Find the maximum value point exceeding the angular velocity detection threshold ω(λ) in the curve as the threshold point. Take the first minimum value point before the threshold point in the curve as the moment TO, and the first minimum value point after the threshold point as the moment HS. If the current gait-related data is in the time interval from the moment TO to the moment HS, it is marked as the stance phase. If the current gait-related data is in the time interval from the moment HS to the moment TO, it is marked as the swing phase.

10. A rehabilitation exoskeleton device, characterized in that, It includes a multi-node wearable data acquisition system, a driving device and a controller. The controller includes the abnormal gait real-time detection system according to any one of claims 1 to 7. The multi-node wearable data acquisition system is used to collect the gait-related data of the patient and transmit it to the controller. The controller is used to input the gait-related data and the pathological data of the patient into the abnormal gait real-time detection system, and control the driving device to assist the corresponding lower limb movement after obtaining the current gait recognition result.

Images