Intelligent wearable rehabilitation monitoring and evaluating device used after anterior cruciate ligament injury

Abstract

The invention provides an intelligent wearable rehabilitation monitoring and evaluation device after anterior cruciate ligament injury, and relates to the technical field of exercise rehabilitation and intelligent medical monitoring, the device comprises a sensor acquisition module, a signal processing module, a compensation determination module, a difference determination module and an evaluation analysis module, a force-electricity coupling flexible sensor, an IMU and a pressure sensor, the method comprises the following steps: synchronously collecting a muscle force signal, a surface electromyogram signal, an attitude signal and a plantar pressure signal at the same position, performing dynamic filtering, confidence enhancement and personalized compensation processing, extracting a difference feature vector, inputting the difference feature vector into a muscle force prediction model, and outputting a quantitative evaluation result. The portable wearable long-term dynamic monitoring is realized, the core indexes of muscles, joints and gaits are covered in multiple dimensions, the assessment is accurate and real-time, the rehabilitation scheme can be optimized, the risk of re-injury is reduced, and the rehabilitation effect and compliance of a patient are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of sports rehabilitation and intelligent medical monitoring, and in particular to an intelligent wearable rehabilitation monitoring and evaluation device after anterior cruciate ligament injury. BACKGROUND

[0002] Anterior cruciate ligament (ACL) injury is a common type of sports injury in the field of sports medicine, and postoperative rehabilitation and return-to-sport decision-making are the core problems of clinical concern. Existing research shows that the proportion of patients returning to sports after ACL reconstruction varies greatly, ranging from 33% to 92%, and the re-injury rate can exceed 20%. The core reason for this phenomenon is that traditional rehabilitation monitoring relies on subjective evaluation scales and periodic hospital equipment detection, lacking continuous and objective quantitative basis. Current rehabilitation monitoring technology has many shortcomings: first, traditional monitoring equipment is bulky and has poor portability, making it impossible to achieve real-time monitoring during daily rehabilitation; second, physiological signal acquisition is easily affected by power frequency interference and motion artifacts, and existing filtering methods are mostly fixed parameter filtering, which cannot be adapted to the biomechanical characteristics of rehabilitation movements, resulting in low signal-to-noise ratio and distortion of core features; third, there is a lack of compensation mechanism for individual differences among patients, and different patients have different individual differences, so a unified monitoring standard cannot accurately reflect the actual functional status.

[0003] With the development of wearable technology and artificial intelligence, wearable sensors have begun to be applied to ACL rehabilitation monitoring for capturing asymmetry, training load and fatigue state during exercise, but existing devices still have problems such as low signal processing accuracy, single evaluation index and poor individual adaptability. At the same time, multi-modal data fusion and machine learning models have been applied in biomechanical analysis, but there is still no special technical solution for ACL rehabilitation monitoring.

[0004] Therefore, the present application provides an intelligent wearable rehabilitation monitoring and evaluation device after anterior cruciate ligament injury. SUMMARY

[0005] The present application provides an intelligent wearable rehabilitation monitoring and evaluation device after anterior cruciate ligament injury to solve the technical problems mentioned above.

[0006] The present application provides an intelligent wearable rehabilitation monitoring and evaluation device after anterior cruciate ligament injury, comprising: A sensor acquisition module is used to wear different types of sensors on the designated positions of the patient, and to acquire N physiological signals for each designated position of the patient, respectively, wherein the physiological signals include force muscle signals, surface electromyography signals, posture signals and plantar pressure signals. The signal processing module is configured to filter, enhance and analog-to-digital convert each physiological signal to obtain a digital signal, and construct a signal time sequence matrix of a corresponding physiological category. The compensation determining module is configured to obtain a posture behavior profile set of each current patient posture, and obtain a behavior difference set of a corresponding specified position in combination with a reference behavior profile set of a corresponding standard patient posture, and perform fitting processing on all behavior difference sets of the same specified position to determine a discrete behavior point and an occurrence frequency of the discrete behavior point as a compensation factor, and the occurrence frequency is less than or equal to N. The difference determining module is configured to adjust a corresponding row vector in the signal time sequence matrix according to the compensation factor and a behavior profile difference set of the row vector to obtain a first time sequence matrix, compare the first time sequence matrix with a standard time sequence matrix constructed based on a standard patient posture of a corresponding index position to obtain a difference matrix, and extract a difference feature vector of the difference matrix. The evaluation analysis module is configured to input the difference feature vector of each physiological signal into a muscle strength prediction model, output a quantitative result of a lower limb function state based on each evaluation index, and compare all quantitative results with a preset quantitative table to obtain a final evaluation result.

[0007] Preferably, the signal processing module comprises: The signal intercepting unit is configured to call start and end intercepting time of a corresponding physiological signal from a preset mapping library based on the current patient posture, construct a time domain gate function adaptive to a feature distribution of the corresponding physiological signal, and perform time axis matching on the corresponding physiological signal and the time domain gate function to intercept a target signal segment. The extraction unit is configured to extract, in real time, a gate-in power frequency interference proportion of a sampling time t of the target signal segment , an active segment proportion , and a cycle stability , calculate a synergic adaptive feature value of the sampling time t, and the calculation formula is: , wherein, is a preset minimum positive parameter; The normalization determining unit is configured to determine a feature distribution dispersion degree of the target signal segment, call a reference dispersion degree of the corresponding physiological signal in the preset mapping library to obtain a time domain gate normalization coefficient K, and the calculation formula is: ; The order determining unit is configured to determine a band-pass filter order of the corresponding physiological signal based on the synergic adaptive feature value and the time domain gate normalization coefficient K, and the calculation formula is: , wherein, a preset reference filter order corresponding to the physiological signal; a rounding function; a model adjusting unit configured to match a filter model consistent with a physiological category of the corresponding physiological signal from a type-model correspondence table, and adjust a notch frequency of the filter model based on a notch frequency of the filter model, and in combination with an order of the corresponding filter model; a filter unit configured to perform a convolution operation on the adjusted model and the time-domain gating function, and in combination with K, normalize and compensate the corresponding physiological signal to output a corresponding filtered signal.

[0008] Preferably, the signal processing module further comprises: a confidence determination unit configured to extract a signal ratio of the corresponding filtered signal , a core feature retention degree , in combination with a cooperative adaptation feature value of the corresponding filtered signal , determine a dynamic confidence of the corresponding filtered signal , and the calculation formula is: wherein, is a reference cooperative adaptation feature value in the preset mapping library matched with a standard patient posture of the corresponding filtered signal; an enhancement unit configured to determine a feature value of an enhanced signal according to the following formula to obtain the enhanced signal; wherein, is a feature value of the enhanced signal at time t; is a feature value of the filtered signal at time t; is a feature value of the reference feature at time t.

[0009] Preferably, the compensation determination module comprises: a set construction unit configured to construct a behavior difference set of each current patient posture at a specified position wherein, is a position difference of an i1th current patient posture corresponding to a specified position based on a j1th posture point; a first fitting unit configured to perform internal fitting processing on each behavior difference set of the same specified position to obtain a first linear coefficient and a first discrete point, wherein the fitting target function is: wherein, , is a slope and an intercept of the first linear coefficient; is a biomechanical contribution weight of the j1th posture point; a first statistical unit configured to count a first occurrence frequency of a first discrete point of a same gesture point based on all behavior difference sets in a same specified position; a second fitting unit configured to perform linear fitting analysis on all the first occurrence frequencies to obtain a second linear coefficient and a second discrete point, and count a second occurrence frequency of the second discrete point of the same gesture point based on all the behavior difference sets in the same specified position; a threshold value determination unit configured to construct a frequency threshold value of the j1th gesture point according to the first occurrence frequency, the second occurrence frequency, the first linear coefficient and the second linear coefficient a factor determination unit configured to filter a third occurrence frequency greater than the corresponding frequency threshold value from all the first occurrence frequencies, and regard the gesture point matched with the third occurrence frequency as a discrete behavior point to obtain a compensation factor wherein, is the third occurrence frequency of the j1th gesture point.

[0010] Preferably, the threshold value determination unit comprises: wherein, is a biomechanics weighted mean of the first occurrence frequency, and , is the first occurrence frequency of the j1th gesture point. is a biomechanics weighted mean of the second occurrence frequency, and , is the second occurrence frequency of the j1th gesture point. is a current rehabilitation stage coefficient of the corresponding patient. is a confidence of the fitting target function. is a determination coefficient of the slope and the intercept based on the second linear coefficient.

[0011] Preferably, the difference determination module comprises: a difference determination unit configured to obtain a time stamp and a gesture identifier corresponding to each row vector in the signal time sequence matrix, and perform double matching of time axis and gesture dimension with the behavior profile difference set of the same specified position to determine a target behavior profile difference set corresponding to each row vector. an element adjustment unit configured to extract an effective discrete behavior point in the compensation factor and a real-time weighted frequency, combine a previous cruciate ligament rehabilitation biomechanics phase weight corresponding to the corresponding behavior point to determine a feature difference degree of each row vector and the corresponding target behavior profile difference set, and obtain an adjustment intensity coefficient to perform adjustment. a recombination unit configured to recombine the adjusted row vectors in the arrangement order of the original time sequence matrix to form a first time sequence matrix. ​​

[0012] Preferably, it further comprises: A vector generation module is configured to generate a global feature vector based on a force-electricity coupling feature grid mining branch, a motion-force dynamic weight coordination branch, and a rehabilitation evaluation score mapping weight layer arranged at the output end of the double branches. A sub-model training module is configured to train and construct exclusive sub-models corresponding to each lower limb function evaluation index based on the global feature vector, as a muscle strength prediction model.

[0013] Preferably, the vector generation module comprises: The force-electricity coupling feature grid mining branch is configured to divide the force-electricity coupling difference feature sequence into fixed-length time sequence grid units, and output a gridized force-electricity coupling feature vector by using a 1D-CNN and a bidirectional LSTM structure, wherein the 1D-CNN layer extracts local space-time correlation features within each grid unit, and the bidirectional LSTM layer captures long-time dependence between grid units. The motion-force dynamic weight coordination branch is configured to output a motion-force coordination feature vector by using a time sequence attention mechanism with a dynamic weight vector and a GRU structure, wherein the time sequence attention mechanism allocates an adaptive weight vector based on physiological features for different phase difference features within a gait cycle, the weight vector is weighted and fused with the output features of the GRU layer, and the dynamic coordination relationship between the posture and the plantar pressure signal is modeled. The rehabilitation evaluation score mapping weight layer is configured to collect the clinical rehabilitation evaluation scores of the patient, map the scores into multi-dimensional weight coefficients, and allocate the fusion weight of the double-branch feature vectors to generate the global feature vector.

[0014] Compared with the prior art, the application has the following beneficial effects: Through the cooperation of multi-modal wearable sensors, hierarchical signal processing, individualized compensation, and the integration of comprehensive evaluation models, the portable, real-time, and multi-dimensional quantitative evaluation of anterior cruciate ligament injury rehabilitation is realized, and the problems of poor portability, single evaluation index, and strong subjectivity of traditional devices are solved, thereby providing a basis for rehabilitation process monitoring and training scheme optimization.

[0015] Other features and advantages of the present application will be further described in the following description, and some will become apparent from the description, or will be learned by practice of the present application. The purposes and other advantages of the present application can be achieved and obtained by the structures specifically pointed out in the written description and drawings.

[0016] The technical solutions of the present application will be further described in detail below with the help of the drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0017] The accompanying drawings are included to provide a further understanding of the application, and are incorporated in and constitute a part of this specification, illustrate embodiments of the application, and together with the description serve to explain the application, and do not limit the application. In the drawings: Figure 1 Figure 1 is a structural diagram of an intelligent wearable rehabilitation monitoring and evaluation device after an anterior cruciate ligament injury according to an embodiment of the application. DETAILED DESCRIPTION

[0018] The preferred embodiments of the application will be described below with reference to the accompanying drawings, and it should be understood that the preferred embodiments described herein are only used to explain and illustrate the application, and do not limit the application.

[0019] The application provides an intelligent wearable rehabilitation monitoring and evaluation device after an anterior cruciate ligament injury, as shown in Figure 1 The device comprises: A sensor acquisition module is configured to wear different types of sensors at designated positions of a patient, and to acquire N physiological signals of each designated position of the patient, wherein the physiological signals include force muscle signals, surface electromyography signals, posture signals, and plantar pressure signals. A signal processing module is configured to filter, enhance, and analog-to-digital convert each physiological signal to obtain a digital signal, and to construct a signal time sequence matrix of the corresponding physiological type. A compensation determination module is configured to obtain a posture behavior profile set of each posture of the patient, and to obtain a behavior difference set of the corresponding designated position in combination with a reference behavior profile set of a standard patient posture, and to determine a discrete behavior point and an occurrence frequency of the discrete behavior point as a compensation factor by fitting all behavior difference sets of the same designated position, wherein the occurrence frequency is less than or equal to N. A difference determination module is configured to adjust a corresponding row vector in the signal time sequence matrix according to the compensation factor and a behavior profile difference set of each row vector to obtain a first time sequence matrix, to compare the first time sequence matrix with a standard time sequence matrix constructed based on a standard patient posture of the corresponding index position to obtain a difference matrix, and to extract a difference feature vector of the difference matrix. An evaluation analysis module is configured to input the difference feature vector of each physiological signal into a muscle strength prediction model, to output a quantitative result of a lower limb function state based on each evaluation index, and to compare all quantitative results with a preset quantitative table to obtain a final evaluation result.

[0020] The specified positions refer to the key positions of the lower extremities of the patient that are closely related to the functional recovery of the anterior cruciate ligament. The physiological signals of these positions can directly reflect the joint stability, muscle strength, and other core rehabilitation indicators. For example, the specified positions for the biomechanical characteristics and clinical monitoring requirements of anterior cruciate ligament rehabilitation include the anterior rectus femoris muscle belly, the medial vastus medialis muscle belly (for muscle force electrical signal collection), the upper edge of the patella 10 cm (lateral thigh), the lateral ankle (for knee joint range of motion collection), the plantar heel area (measurement point 1), and the first and second metatarsal interosseous area (measurement point 2, for plantar pressure collection). These positions have been verified through clinical experiments and can capture the physiological signals related to rehabilitation to the greatest extent.

[0021] The different types of sensors refer to special wearable sensors that adapt to the collection of four types of physiological signals. They need to meet the characteristics of small size, good flexibility, strong endurance, and high sampling accuracy. For example, three types of sensors are used, including a force-electricity coupling flexible sensor (five-layer structure, PET material for the base layer, P(VDF-HFP) based ion gel for the pressure sensitive layer, and silver paste printed electrode for the signal transmission layer) for synchronous collection of surface electromyography signals and muscle deformation mechanics signals, a 6-axis inertial measurement unit (IMU, model MPU6050, containing a three-axis accelerometer and a three-axis gyroscope) for collection of knee joint kinematics signals, and a flexible force-sensitive resistor (FSR) pressure sensor (PET for the base layer and PEDOT:PSS material for the pressure sensitive layer) for collection of plantar pressure signals. These three types of sensors meet the requirements of wearable design and are fixed at the specified positions by elastic bands.

[0022] N times of collection refers to multiple repeated collections to reduce the influence of accidental factors on signal quality and ensure data reliability. The collection process needs to control the consistency of the environment and actions. For example, N is 5, each collection lasts for 30 seconds, and the interval between two collections is 1 minute (to avoid muscle fatigue). The patient needs to perform standard actions (such as straight leg lifting, knee flexion and extension, and walking on flat ground) in a flat environment. The collection device synchronously records the start and end time stamps of the actions to ensure the accurate alignment of the signal and action time axis.

[0023] The four types of physiological signals refer to the core physiological signals that can comprehensively reflect the functional status of the lower extremities, including force muscle signals, surface electromyography signals, posture signals, and plantar pressure signals. Force muscle signals: mechanical pressure signals generated when muscles contract, reflecting muscle contraction intensity and force efficiency. For example, the pressure changes of the rectus femoris muscle contraction are collected by a force-electricity coupling flexible sensor, with a signal amplitude range of 0-5V and a sampling rate of 1600Hz, which can capture the dynamic changes of muscle pressure from relaxation to maximum contraction.

[0024] Surface electromyography signal: bioelectric signal generated by muscle movement, its amplitude and frequency characteristics are related to muscle activity state and fatigue degree. For example, the signal amplitude range is -1000 ~1000 , the sampling rate is 1600 Hz, and the median frequency and average power frequency can be extracted to reflect muscle fatigue.

[0025] Posture signal: contains joint angle, angular velocity, acceleration and other parameters, reflecting joint range of motion and motion control ability. For example, the angle change (range 0°-120°) and angular velocity (range 0-50° / s) of the knee joint during flexion and extension are collected by IMU, with a sampling rate of 100 Hz, which can accurately capture the joint motion trajectory.

[0026] Plantar pressure signal is the interaction force signal between the plantar and the support surface during gait, reflecting the gait symmetry and force distribution. For example, the pressure value (range 0-100 N) between the heel and the first and second metatarsal bones is collected by the plantar FSR sensor, with a sampling rate of 100 Hz, which can record the time-varying curve of pressure in the gait cycle.

[0027] Signal filtering, enhancement and analog-to-digital conversion refer to the process of noise removal, feature enhancement and analog-to-digital conversion of original physiological signals to ensure that the signal quality meets the subsequent analysis requirements.

[0028] Signal time sequence matrix refers to the matrix formed by arranging the digital signals of the same physiological category in chronological order, with each row corresponding to the feature data at a time point and each column corresponding to a feature type. For example, taking surface electromyography signal as an example, 5 times of 30 seconds are collected, with a sampling rate of 1600 Hz, and 48000 data points (30s x 1600 Hz) are obtained each time. The signal time sequence matrix is 5 rows and 48000 columns, each row corresponds to the time sequence data of one collection, and the element is the electromyography signal amplitude at that time point.

[0029] Posture behavior profile set refers to the set of posture parameters at specified positions changing with time when the patient performs rehabilitation movements, including joint angle, position coordinates and other data at each time. For example, when the patient performs knee flexion and extension movements, the angle data collected by the IMU at 10 cm above the patella changes with time, such as the sequence of angle from 0° gradually increasing to 90° and then restoring to 0° within 0-30 seconds, which is the posture behavior profile set of the movement.

[0030] The reference behavior profile set refers to a standard posture template constructed based on a large number of healthy people (20-40 years old, no history of knee joint injury) and rehabilitation standard patients. For example, for the knee joint flexion and extension action, the reference behavior profile set is a uniform speed change curve from 0° to 90°, the average angular velocity is 30° / s, and the angle change standard deviation is ≤5°. This template is obtained by averaging the measured data of 100 healthy people.

[0031] The behavior difference set refers to the difference set of the real-time collected posture behavior profile set and the reference behavior profile set, which is used to quantify the deviation of the actual action from the standard action. For example, in a certain knee joint flexion and extension action of a patient, the real-time angle at the 10th second is 45°, and the angle in the reference behavior profile set at this time is 40°. Therefore, the angle difference at this time is 5°, and the behavior difference set is a set composed of all such deviation data at all times, such as {2°, 3°, 5°,..., 4°}.

[0032] The fitting processing refers to analyzing the behavior difference set using mathematical fitting methods to identify discrete points deviating from the overall trend. For example, linear fitting is used to process the behavior difference set, and the fitting target function is to minimize the weighted sum of squares (the weight is the biomechanical contribution weight of the posture point), and the fitting straight line equation y=k1x+b1 (k1 is the slope and b1 is the intercept). The points deviating from the straight line and having a deviation absolute value greater than 3° are the first discrete points.

[0033] The discrete behavior point refers to the posture point deviating from the overall trend after fitting processing, which is related to the patient's action being non-standard or the functional recovery being incomplete. For example, in the behavior difference set of the knee joint flexion and extension, the fitting straight line is y=0.2x+1.5, and the deviation value at a certain time is 8°, which is much higher than the deviations at other times (2-5°). The posture point corresponding to this time is the discrete behavior point.

[0034] The compensation factor refers to the set of discrete behavior points and their occurrence frequencies, which is used to adjust the subsequent signal time sequence matrix, wherein the occurrence frequency ≤N (collection times). For example, a certain discrete behavior point (corresponding to the knee joint flexion to 60°) appears 3 times in 5 collections, and its occurrence frequency 3 ≤5. Therefore, the compensation factor is {3, 60° posture point}.

[0035] The first time sequence matrix refers to the optimized matrix formed by adjusting the row vectors of the original signal time sequence matrix according to the compensation factor, which can more accurately reflect the actual functional status of the patient. For example, in the original signal time sequence matrix, the signal amplitude of a certain row vector (corresponding to the 3rd collection) at the timestamp of the discrete behavior point is 3.2V. According to the compensation factor (the discrete point appears 3 times, and the biomechanical weight is 0.3), the adjustment intensity coefficient is calculated as 0.8, and the adjusted amplitude is 3.2V×0.8=2.56V. Arranging all the adjusted row vectors in the original order obtains the first time sequence matrix.

[0036] The standard timing matrix refers to a timing matrix constructed based on the physiological signals corresponding to the reference behavior profile set as a benchmark for evaluation. For example, using the surface electromyography signals corresponding to the reference behavior profile set, a timing matrix is constructed according to the same acquisition parameters (5 acquisitions, 30 seconds each, sampling rate 1600 Hz), and the elements of the timing matrix are the standard amplitudes of the signals of healthy people.

[0037] The difference matrix refers to a matrix composed of the difference between the corresponding elements of the first timing matrix and the standard timing matrix, which is used to reflect the overall difference between the real-time signal and the standard signal. For example, if an element in the first timing matrix is 2.56V and the corresponding element in the standard timing matrix is 2.8V, then the element in the difference matrix is -0.24V, and all the elements in the difference matrix collectively reflect the deviation distribution of the signal.

[0038] The difference feature vector refers to a combination of features extracted from the difference matrix that can characterize the core difference, covering multiple dimensions such as amplitude, frequency, and time domain. For example, for the difference matrix of surface electromyography signals, the extracted difference feature vector is: {maximum deviation (0.5V), average deviation (0.2V), median frequency deviation (2Hz), average power frequency deviation (3Hz)}, which can comprehensively reflect the difference between the real-time signal and the standard signal.

[0039] The muscle strength prediction model refers to a model constructed based on machine learning algorithms, which can output the quantitative results of lower limb function evaluation indicators when the difference feature vector is input.

[0040] The preset quantification table refers to an evaluation threshold table developed in combination with clinical rehabilitation standards, which is used to convert the quantitative results into rehabilitation effect grades. For example, the preset quantification table contains 4 evaluation indicators, each with a maximum score of 100 points, and the specific thresholds are as follows: muscle strength level ≥ 80 points for good rehabilitation, 60-79 points for moderate rehabilitation, and < 60 points for insufficient rehabilitation; gait symmetry ≥ 85 points for good rehabilitation, 70-84 points for moderate rehabilitation, and < 70 points for insufficient rehabilitation; joint range of motion ≥ 90 points for good rehabilitation, 75-89 points for moderate rehabilitation, and < 75 points for insufficient rehabilitation; muscle fatigue degree ≤ 30 points for good rehabilitation, 31-50 points for moderate rehabilitation, and > 50 points for insufficient rehabilitation. The final evaluation result is the average score of the 4 indicators.

[0041] The beneficial effects of the above technical solutions are: through the integration of multi-modal wearable sensor cooperative acquisition, hierarchical signal processing, individualized compensation, and comprehensive evaluation model, the portable, real-time, and multi-dimensional quantitative evaluation of anterior cruciate ligament injury rehabilitation is realized, solving the problems of poor portability, single evaluation index, and strong subjectivity of traditional devices, providing a basis for rehabilitation progress monitoring and training scheme optimization.

[0042] This invention provides an intelligent wearable rehabilitation monitoring and assessment device for anterior cruciate ligament injury, wherein the signal processing module includes: The signal interception unit is used to call the start and end interception times of the corresponding physiological signal from the preset mapping library based on the patient's posture at that time, construct a time-domain gate function that is adapted to the feature distribution of the corresponding physiological signal, and match the corresponding physiological signal with the time-domain gate function on the time axis to intercept the target signal segment. The extraction unit is used to extract the in-gate power frequency interference ratio at sampling time t of the target signal segment in real time. Activity segment ratio and periodic stability Calculate the cooperative adaptation feature value at acquisition time t. The calculation formula is: ,in, The preset minimum positive parameter; The normalization determination unit is used to determine the feature distribution dispersion of the target signal segment. The baseline dispersion of the corresponding physiological signal in the preset mapping library is called. The time-domain gate normalization coefficient K is obtained by the following formula: ; The order determination unit is used to determine the order based on the cooperative adaptation feature value. The time-domain gate normalization coefficient K determines the bandpass filter order corresponding to the physiological signal. The calculation formula is: ,in, This is the preset benchmark filter order corresponding to the physiological signal; This is a rounding function; The model adjustment unit is used to match a filtering model from the type-model lookup table that corresponds to the physiological type of the corresponding physiological signal, and based on... The notch frequency of the filtering model is dynamically adjusted, and combined with... Adjust the order of the corresponding filtering model; The filtering unit is used to perform convolution operations between the adjusted model and the time-domain gate function, and to normalize and compensate the corresponding physiological signal in combination with K, and output the corresponding filtered signal.

[0043] In this embodiment, the preset mapping library refers to a database storing baseline parameters corresponding to different rehabilitation movements and physiological signal types. It includes parameters such as start and end times, baseline dispersion, and baseline co-adaptation feature values, and can be constructed using clinical trial data. For example, the preset mapping library is a locally stored JSON file, where the start and end times of the surface electromyography signal corresponding to the straight leg raise movement are 0.5 to 3 seconds after the start of the movement, and the baseline dispersion... =0.8, benchmark co-fit eigenvalue =0.7; the start and end time of the posture signal corresponding to the knee flexion and extension action is 0.3 seconds to 2.8 seconds after the action starts, the reference dispersion =0.6, the reference collaborative fitting characteristic value =0.65, and the device local chip is called.

[0044] In this embodiment, the time domain gate function refers to a time domain screening function for fitting the physiological signal feature distribution, which is used to intercept the effective signal segment and remove the invalid signal before and after the action starts. For example, a rectangular window function is used as the time domain gate function. For the surface electromyography signal of the straight leg raising action, the time range of the time domain gate function is 0.5 seconds to 3 seconds (consistent with the preset mapping library), the window width is 2.5 seconds, the signal amplitude in the window is retained, the signal amplitude outside the window is set to 0, and the effective signal segment is intercepted.

[0045] The power frequency interference proportion in the gate refers to the proportion of the energy of the 50Hz power frequency interference signal in the target signal segment to the total energy of the target signal segment, which reflects the severity of the power frequency interference. For example, the frequency spectrum of the target signal segment is analyzed by using the Fourier transform, the signal energy at the 50Hz frequency is 0.05mJ, and the total energy of the target signal segment is 0.5mJ. Then =0.05 / 0.5=0.1.

[0046] The activity segment proportion refers to the proportion of the time length of the muscle contraction activity period (the period when the signal amplitude is higher than the threshold value) in the target signal segment to the total time length of the target signal segment, which reflects the duration of muscle activity. For example, the time length of the target signal segment of the straight leg raising action is 2.5 seconds, and the time length of the muscle contraction activity period (the electromyography signal amplitude > 50μV) is 1.7 seconds. Then =1.7 / 2.5=0.68.

[0047] The cycle stability refers to the consistency of the repetition cycle of the rehabilitation action in the target signal segment, which is obtained by calculating the ratio of the standard deviation to the mean value of each cycle. For example, the straight leg raising action is repeated 3 times in the target signal segment, the cycle time lengths are 0.8 seconds, 0.85 seconds and 0.75 seconds respectively, the cycle mean value is 0.8 seconds, and the standard deviation is 0.05 seconds. Then =0.05 / 0.8=0.0625.

[0048] In this embodiment, the preset minimum positive parameter is used to avoid the denominator being 0.

[0049] The feature distribution dispersion refers to the dispersion degree of the feature data (such as amplitude) of the target signal segment, which is obtained by calculating the variance of the feature data. For example, the surface electromyography signal amplitude data of the target signal segment is , and the variance calculated is the dispersion.

[0050] In this embodiment, the preset reference filter order The initial filter order corresponding to each physiological signal is preset, which is determined based on signal characteristics and clinical experiments and stored in the type-model correspondence table, as shown in Table 1: Table 1 Type-model correspondence table The notch frequency refers to a frequency parameter in the filter model for suppressing specific frequency interference (such as power frequency interference), which can be dynamically adjusted based on the proportion of power frequency interference in the door. For example, the basic notch frequency of the surface electromyography signal is 50 Hz, when =0.1 (light interference), the notch frequency is adjusted to 50 Hz; when =0.3 (heavy interference), the notch frequency is adjusted to a frequency band of 49.5-50.5 Hz.

[0051] The convolution operation refers to mathematical convolution of the adjusted filter model and the time domain gate function, which realizes precise filtering of the target signal segment, removes noise while retaining effective features. For example, the adaptive notch filter model with an order of 4 is convolved with a rectangular time domain gate function (window width 2.5 seconds), the convolution kernel size is 3, the step is 1, and the target signal segment of the surface electromyography signal is operated point by point to obtain the filtered signal.

[0052] The beneficial effects of the above technical solutions are: by dynamically adjusting the filter parameters (order, notch frequency), precise filtering of physiological signals is realized, which adapts to the signal processing needs under different actions and different interference levels, effectively removes power frequency interference, motion artifacts and other noise, and at the same time, the effective signal segment is intercepted through the time domain gate function, which improves the pertinence and efficiency of signal processing.

[0053] The present application provides a kind of intelligent wearable rehabilitation monitoring and evaluation device after anterior cruciate ligament injury, the signal processing module further includes: confidence determination unit, for extracting the signal ratio , core feature retention , combined with the cooperative adaptation characteristic value of corresponding filtered signal, determines the dynamic confidence of corresponding filtered signal, and the calculation formula is as follows: , wherein, is the reference cooperative adaptation characteristic value matched with the standard patient posture of corresponding filtered signal in the preset mapping library; enhancement unit, the characteristic value of enhanced signal is determined according to the following formula, and enhanced signal is obtained; , wherein, is the characteristic value of enhanced signal at time t. is the eigenvalue of the filtered signal at time t; is the eigenvalue of the reference feature at time t.

[0054] In this embodiment, the signal ratio is the ratio of the effective signal amplitude to the noise signal amplitude of the filtered signal, reflecting the quality of the filtered signal, with the unit of dB. For example, the effective signal amplitude (muscle contraction period) of the filtered surface electromyography signal is 200 , and the noise signal amplitude (muscle relaxation period) is 10 , then = 20lg(200 / 10) = 26dB.

[0055] Core feature retention degree is the ratio of the core feature (such as the peak value of the electromyography signal or the maximum pressure value of the force myography signal) of the filtered signal to the core feature of the original signal, reflecting the retention degree of the core information in the filtering process. For example, the peak value of the original surface electromyography signal is 300 , and the peak value of the filtered signal is 285 , then the core feature retention degree = 285 / 300 = 0.95.

[0056] In this embodiment, the eigenvalue of the enhanced signal refers to the optimized eigenvalue obtained by dynamically weighting the eigenvalue of the filtered signal and the eigenvalue of the reference feature, which can improve the reliability and recognition of the signal.

[0057] The beneficial effects of the above technical solution are: signal enhancement is realized through dynamic confidence weighting, which not only retains the real-time features of the filtered signal, but also improves the reliability and stability of the signal through calibration of the reference eigenvalue, effectively solving the problem that a single signal is easily affected by individual differences and environmental interference, and the enhanced signal can more accurately reflect the physiological state of the patient.

[0058] The application provides an intelligent wearable rehabilitation monitoring and evaluation device after an anterior cruciate ligament injury. The set construction unit is configured to construct a behavior difference set of each current patient posture at a specified position , wherein is the position difference of the ith1 current patient posture corresponding to the specified position based on the jth1 posture point; The first fitting unit is configured to perform internal fitting processing on each behavior difference set of the same specified position to obtain a first linear coefficient and a first discrete point, wherein the fitting target function is: , wherein , a slope of the first linear coefficient, an intercept; a biomechanics contribution weight of the j1th posture point; a first statistical unit configured to count a first occurrence frequency of the first discrete point of the j1th posture point based on all the behavior difference sets in the same specified position; a second fitting unit configured to perform linear fitting analysis on all the first occurrence frequencies to obtain a second linear coefficient and a second discrete point, and count a second occurrence frequency of the second discrete point of the j1th posture point based on all the behavior difference sets in the same specified position; a threshold determining unit configured to construct a frequency threshold of the j1th posture point according to the first occurrence frequency, the second occurrence frequency, the first linear coefficient and the second linear coefficient ; a factor determining unit configured to screen a third occurrence frequency greater than the corresponding frequency threshold from all the first occurrence frequencies, and regard the posture point matched with the third occurrence frequency as a discrete behavior point to obtain a compensation factor , wherein, the third occurrence frequency of the j1th posture point.

[0059] Preferably, the threshold determining unit comprises: ; wherein, the biomechanics weighted mean of the first occurrence frequency, and , the first occurrence frequency of the j1th posture point; the biomechanics weighted mean of the second occurrence frequency, and , the second occurrence frequency of the j1th posture point; a current rehabilitation stage coefficient of the corresponding patient; a confidence degree of the fitting target function; a determination coefficient of the slope and the intercept based on the second linear coefficient.

[0060] In this embodiment, the behavior difference set refers to a behavior difference set corresponding to the i1th posture of the patient at the moment, which contains the position difference of each posture point in the posture. For example, i1=3 (the 3rd straight leg lifting action), and the posture points j1=1, 2,..., 10 (corresponding to 10 key time points of the action, with an interval of 0.25 seconds). The position differences of the posture points are respectively: 2°, 3°, 5°, 4°, 6°, 3°, 2°, 4°, 3°, 2°, which further constitute the behavior difference set .

[0061] ​​The posture point j1 refers to a key node of dividing the posture behavior profile set of the rehabilitation action in the time or space dimension, and each posture point corresponds to a specific action state. For example, the 30-second posture behavior profile set of the straight leg lifting action is evenly divided into 10 posture points (j1 = 1 to 10), j1 = 1 corresponds to 0.5 seconds after the action starts, j1 = 5 corresponds to the middle stage (1.5 seconds), and j1 = 10 corresponds to 0.25 seconds before the action ends, and each posture point corresponds to a knee joint angle value.

[0062] The first linear coefficient k1 and b1 refer to the slope and intercept of the fitting straight line obtained by linear fitting of the behavior difference set, reflecting the overall trend of the behavior difference. For example, the behavior difference set is linearly fitted, and k1 = 0.1 and b1 = 2.8 are calculated, and the fitting straight line equation is y = 0.1x + 2.8.

[0063] Biomechanical contribution weight refers to the weight allocated according to the biomechanical importance of each posture point in the rehabilitation action, and the important posture point has a higher weight. For example, in the straight leg lifting action, the biomechanical contribution weight of j1 = 5 (the muscle force is the largest in the middle stage of the action) is = 0.2, the weight of j1 = 1 and 10 (the start and end stages of the action) is = 0.08, and the weight of the remaining posture points is 0.1. The sum of all weights is 1.0, and the weight is determined through biomechanical simulation and clinical experiment.

[0064] The first discrete point refers to the point in the behavior difference set that deviates from the first linear fitting straight line, and the deviation degree is judged according to the preset threshold value (such as the absolute value of the deviation > 3°). For example, the fitting straight line equation is y = 0.1x + 2.8, and when j1 = 5 = 6°, the fitting value is 0.1x5 + 2.8 = 3.3°, the absolute value of the deviation is 6-3.3 = 2.7° < 3°, and it is not a first discrete point; when j1 = 4 = 4°, the fitting value is 0.1x4 + 2.8 = 3.2°, the deviation is 0.8°, and it is not a discrete point; when j1 = 6 = 3°, the fitting value is 0.1x6 + 2.8 = 3.4°, the deviation is 0.4°, and there is no first discrete point; if j1 = 3 in a certain behavior difference set = 9°, the fitting value is 3.1°, and the deviation is 5.9° > 3°, so the point is a first discrete point.

[0065] The first occurrence frequency refers to the number of times the first discrete point appears in the behavior difference set of all N times of collection at the same specified position and the same posture point. For example, N = 5 times of collection, and the posture point j1 = 3 appears as a first discrete point in the 1st, 3rd, and 5th times of collection, and does not appear in the 2nd and 4th times of collection, so the first occurrence frequency F1,3 = 3. ​

[0066] The second linear coefficient k2, b2 refers to the slope and intercept of the fitting straight line obtained by linear fitting of the first occurrence frequency of all posture points, reflecting the overall trend of the first occurrence frequency. For example, the first occurrence frequency of each posture point in this embodiment is {1, 2, 3, 1, 2, 1, 0, 1, 2, 1}, and the linear fitting obtains k2 = 0.05 and b2 = 1.2, and the fitting straight line equation is y = 0.05x + 1.2.

[0067] The second discrete point refers to the point deviating from the second linear fitting straight line in the first occurrence frequency, and the deviation degree is judged according to the preset threshold (such as the absolute value of the deviation > 1). For example, the first occurrence frequency F1,3 of j1 = 3 is 3, the fitting value is 0.05x3 + 1.2 = 1.35, and the deviation is 3-1.35 = 1.65 > 1, so the point is the second discrete point.

[0068] The second occurrence frequency refers to the number of times the second discrete point appears in the behavior difference set of the same specified position and the same posture point in all N times of collection. For example, the posture point j1 = 3 appears the second discrete point twice in 5 times of collection, and the second occurrence frequency = 2.

[0069] In this embodiment, the current rehabilitation stage coefficient refers to the coefficient set according to the current rehabilitation stage of the patient, which is different for different rehabilitation stages, reflecting the influence of the rehabilitation process on the discrete behavior point. For example, the rehabilitation stage is divided into early stage (1-3 weeks after operation), middle stage (4-8 weeks after operation), and late stage (9-12 weeks after operation), and the corresponding is 0.8, 1.0 and 1.2 respectively. In this embodiment, the patient is in the middle stage of rehabilitation, = 1.0.

[0070] The confidence C1 of the fitting target function refers to the confidence coefficient of the first linear fitting, reflecting the fitting degree of the fitting straight line and the behavior difference set, and the value range is 0-1, and the closer to 1, the better the fitting effect. For example, the first linear fitting 0.78, and the confidence C1 of the fitting target function is 0.78.

[0071] The determination coefficient refers to the fitting degree based on the slope k2 and intercept b2 of the second linear coefficient, reflecting the effect of the second linear fitting, and the value range is 0-1. For example, the second linear fitting = 0.78, indicating that 78% of the first occurrence frequency can be explained by the second fitting straight line.

[0072] ​The beneficial effects of the above technical solutions are: through twice linear fitting and frequency statistics, the discrete behavior points reflecting patient individual differences or functional recovery deviation are accurately identified, the targeted compensation factor is formed, a scientific basis is provided for subsequent signal adjustment, the influence of accidental errors on the evaluation result is effectively reduced, and the accuracy of the evaluation is improved. By comprehensively considering the overall level of the first and second occurrence frequencies, the fitting effect, the biomechanical weight and the rehabilitation stage, the individualization and accuracy of the frequency threshold are ensured, an objective basis is provided for the screening of effective discrete behavior points, and misjudgment or omission caused by a fixed threshold is avoided.

[0073] The application provides an intelligent wearable rehabilitation monitoring and evaluation device after an anterior cruciate ligament injury. The difference determination unit is used for acquiring a timestamp and a posture identifier corresponding to each row vector in the signal time sequence matrix, performing double matching of a time axis and a posture dimension with a behavior profile difference set of a same specified position, and determining a target behavior profile difference set corresponding to each row vector. The element adjustment unit is used for extracting an effective discrete behavior point in the compensation factor and a real-time weighted frequency, combining a pre-cruciate ligament rehabilitation biomechanical phase weight corresponding to the corresponding behavior point, determining a feature difference degree of each row vector and the corresponding target behavior profile difference set, obtaining an adjustment intensity coefficient for adjustment, and adjusting the signal time sequence matrix. The recombination unit is used for recombining the adjusted row vectors according to the arrangement order of the original time sequence matrix to form a first time sequence matrix.

[0074] In the embodiment, the timestamp and the posture identifier are a specific time (accurate to milliseconds) of signal collection, and the posture identifier is a rehabilitation action type and a collection number corresponding to the signal, which are used to realize matching of the signal and the behavior difference set. For example, the timestamp corresponding to a certain row vector is T01, and the posture identifier is straight leg lifting-the 3rd time, which indicates that the row vector is signal data at 2.5 seconds in the 3rd straight leg lifting action.

[0075] The double matching refers to time axis matching (the timestamp is consistent) and posture dimension matching (the posture identifier is consistent) of the row vector of the signal time sequence matrix and the behavior difference set, so as to ensure the accuracy of the corresponding relationship. For example, the time axis matching requires that the timestamp error of the row vector and the timestamp of the behavior difference set is less than or equal to 10 milliseconds, and the posture dimension matching requires that the posture identifier of the row vector and the posture identifier of the behavior difference set are completely consistent (such as straight leg lifting-the 3rd time), and after the double matching, the behavior difference set is determined as the target behavior profile difference set of the column vector.

[0076] The target behavior contour difference set refers to the behavior difference set corresponding to a row vector in the signal time series matrix through double matching, and is used as the basis for adjusting that column vector. For example, if the timestamp of a row vector in the signal time series matrix is ​​2.5 seconds and the posture is identified as straight leg raise - 3rd time, then the difference data corresponding to 2.5 seconds in the behavior difference set of the 3rd straight leg raise action constitutes the target behavior contour difference set, such as {4°}.

[0077] Effective discrete behavior points refer to the discrete behavior points included in the compensation factor, that is, the posture point corresponding to the third occurrence frequency. For example, if the compensation factor is {(3,j1=3)}, then the posture point corresponding to j1=3 (when the straight leg raise action is at 1.0 second) is an effective discrete behavior point.

[0078] Real-time weighted frequency refers to the value of the third occurrence frequency of an effective discrete behavior point, weighted by its biomechanical contribution weight, reflecting the importance of that discrete behavior point. For example, the third occurrence frequency of the effective discrete behavior point j1=3... =3, biomechanical contribution weight =0.2, then the real-time weighted frequency = 3 × 0.2 = 0.6.

[0079] Biomechanical phase weighting in anterior cruciate ligament (ACL) rehabilitation refers to the weighting assigned to the biomechanical phases (such as muscle activation and relaxation phases) of rehabilitation movements. Different phases have different importance in rehabilitation assessment. For example, the biomechanical phases of a straight leg raise are divided into an initiation phase (0.5-1.0 seconds), an activation phase (1.0-2.0 seconds), and a relaxation phase (2.0-3.0 seconds), with corresponding phase weights of 0.3, 0.5, and 0.2, respectively. The effective discrete behavior point j1=3 is in the activation phase, with a phase weight of 0.5.

[0080] Feature difference refers to the degree of difference between the eigenvalues ​​of the row vectors of the signal time-series matrix and the target behavior contour difference set, and is calculated using Euclidean distance. For example, if the eigenvalue of the row vector is 3.2V (muscle signal amplitude) and the difference value of the target behavior contour difference set is 4° (angle deviation), the Euclidean distance after normalization is 0.35, i.e., feature difference = 0.35.

[0081] The adjustment intensity coefficient is a coefficient calculated based on real-time weighted frequency, biomechanical phase weight, and feature difference degree. It is used to determine the adjustment magnitude of the row vector and takes a value between 0 and 1. For example, if real-time weighted frequency = 0.6, biomechanical phase weight = 0.5, and feature difference degree = 0.35, then the adjustment intensity coefficient = 0.6 × 0.5 × 0.35 = 0.105.

[0082] The first time sequence matrix refers to a matrix formed by recombining all the adjusted row vectors in the arrangement order of the original signal time sequence matrix, and is an optimized signal time sequence matrix.

[0083] The above technical solution has the beneficial effects that: the double matching ensures accurate correspondence of the signal and the behavior difference set, the adjustment intensity coefficient is calculated based on the effective discrete behavior points, biomechanics phase weight and feature difference degree, the individual adjustment of the signal time sequence matrix is realized, and the reliability and pertinence of the signal data are improved, thereby laying a high-quality data foundation for subsequent difference comparison and evaluation analysis.

[0084] The application provides an intelligent wearable rehabilitation monitoring and evaluation device after an anterior cruciate ligament injury, and further comprises: The vector generation module is configured to generate a global feature vector based on force-electricity coupling feature gridding mining branches, motion-stress dynamic weight cooperative branches, and a rehabilitation evaluation score mapping weight layer arranged at the output end of the double branches. The sub-model training module is configured to train and construct exclusive sub-models corresponding to each lower limb function evaluation index based on the global feature vector, as a muscle strength prediction model.

[0085] Preferably, the vector generation module comprises: The force-electricity coupling feature gridding mining branch is configured to divide the force-electricity coupling difference feature sequence into time sequence grid units of a fixed length, and output a gridded force-electricity coupling feature vector using a 1D-CNN and a bidirectional LSTM structure, wherein the 1D-CNN layer extracts local space-time correlation features within each grid unit, and the bidirectional LSTM layer captures long-time dependence between grid units. The motion-stress dynamic weight cooperative branch is configured to output a motion-stress cooperative feature vector using a time sequence attention mechanism with a dynamic weight vector and a GRU structure, wherein the time sequence attention mechanism allocates an adaptive weight vector based on physiological features to different phase difference features within a gait cycle, and the weight vector is weighted and fused with the output features of the GRU layer to model the dynamic cooperative relationship between the posture and the plantar pressure signal. The rehabilitation evaluation score mapping weight layer is configured to collect patient clinical rehabilitation evaluation scores, map the scores into multi-dimensional weight coefficients, and allocate fusion weights of the double-branch feature vectors to generate a global feature vector.

[0086] In this embodiment, the exclusive sub-model comprises: The muscle strength level sub-model is used for dividing the measured muscle strength value into a plurality of numerical interval grids, searching for exclusive hyperparameters for each interval grid through Bayesian optimization combined with 5-fold cross-validation, and integrating local models of each grid to obtain a global muscle strength prediction model. The gait symmetry / joint range of motion sub-model: based on gait cycle position coding, the input features are divided into a plurality of feature blocks according to the gait phase, each feature block is mapped into an exclusive feature vector and then input into the Transformer, and the feature contribution degree is adjusted through a dynamic weight vector at the output end of the Transformer to strengthen the phase correlation of the time sequence features. The muscle fatigue degree sub-model is used for dividing the surface electromyography difference features into a plurality of time sequence blocks, calculating the median frequency and average power frequency entropy value of each block, constructing a block-by-block fatigue representation vector, inputting a random forest regression model for training, and realizing fine-grained evaluation of the fatigue state.

[0087] In this embodiment, the branch includes a 1D-CNN layer and a bidirectional LSTM layer, the 1D-CNN layer is used to extract local features of each grid unit, and the bidirectional LSTM layer is used to capture the time sequence correlation between grid units, and output a grid force-electricity coupling feature vector.

[0088] In this embodiment, the time sequence attention mechanism allocates adaptive weights according to the importance of different phases of the gait cycle, and the GRU structure is used to capture the time sequence dependence relationship, and output a motion-force coordination feature vector.

[0089] In this embodiment, the clinical rehabilitation evaluation score is the Lysholm knee score (full score 100), which is mapped into 4-dimensional weight coefficients (corresponding to 4 evaluation indexes) for fusing the grid force-electricity coupling feature vector and the motion-force coordination feature vector.

[0090] In this embodiment, the grid force-electricity coupling feature vector has a dimension of 64, the motion-force coordination feature vector has a dimension of 64, and the fusion weight coefficient is [0.5, 0.5], so the global feature vector is 128, which includes comprehensive features of force, electricity, motion and force.

[0091] The exclusive sub-model refers to a machine learning model trained separately for each lower limb function evaluation index (muscle strength level, gait symmetry, joint range of motion, muscle fatigue degree), which can accurately output the quantitative results of the corresponding index, such as the muscle strength level exclusive sub-model using XGBoost regression, the gait symmetry exclusive sub-model using random forest classification, the joint range of motion exclusive sub-model using support vector regression, and the muscle fatigue degree exclusive sub-model using logistic regression.

[0092] The muscle strength prediction model refers to a model set composed of all the specific sub-models, and the quantification results of all the evaluation indexes can be output by inputting the global feature vector. For example, the muscle strength prediction model includes 4 specific sub-models, and the 128-dimensional global feature vector is input to output 4 quantification results of muscle strength level (such as 85 points), gait symmetry (such as 90 points), joint range of motion (such as 88 points), and muscle fatigue degree (such as 25 points).

[0093] In this embodiment, the time sequence grid unit refers to a unit divided by the force-electric coupling difference feature sequence in a fixed time length, and each unit contains feature data in the time period, facilitating local feature extraction. For example, the force-electric coupling difference feature sequence is collected for 30 seconds, the sampling rate is 1600 Hz, the time sequence grid unit is divided by 0.1 second, each unit contains 160 data points (0.1s x 1600 Hz), and 30 seconds are divided into 300 time sequence grid units.

[0094] In this embodiment, the 1D-CNN layer includes 3 convolution kernels, the convolution kernel size is 3, the step is 1, the activation function is ReLU, and the output feature map dimension is 3 x (300-3+1) = 3 x 298; the bidirectional LSTM layer includes 64 hidden units, the forward LSTM captures the dependency relationship from the 1st to the 300th unit, the backward LSTM captures the dependency relationship from the 300th to the 1st unit, and the output dimension is 64 x 2 = 128, and a 64-dimensional grid force-electric coupling feature vector is obtained after pooling.

[0095] The local spatiotemporal correlation feature refers to the spatial distribution and temporal variation characteristics of the feature data in each time sequence grid unit, which can reflect the local characteristics of the physiological signal in the time period. For example, the force-electric coupling difference feature data in a certain time sequence grid unit (0.1 second) is {2.1V, 2.3V,..., 2.2V}, and the local spatiotemporal correlation feature includes the mean (2.2V), standard deviation (0.08V), and peak value (2.3V) of the data in the unit, which are extracted by the 1D-CNN layer.

[0096] The long-time dependency relationship refers to the feature correlation between different time sequence grid units, reflecting the change trend of the physiological signal in a long time range. For example, in the straight leg lifting action, the force-electric coupling features of the 10th-20th time sequence grid unit (1-2 seconds) gradually increase (muscle strength increases), and the 20th-30th unit (2-3 seconds) gradually decreases (muscle relaxation), and this cross-unit change trend is the long-time dependency relationship, which is captured by the bidirectional LSTM layer.

[0097] The time sequence attention mechanism of the dynamic weight vector refers to dynamically allocating the attention mechanism of the weight vector according to the importance of physiological characteristics in different phases of the gait cycle, and the weight vector is adaptively adjusted with the change of the phase. For example, the gait cycle is divided into a support phase (0-60%) and a swing phase (60-100%), and the cooperativity of the plantar pressure signal and the posture signal in the support phase is more important for evaluation, and the weight vector [0.7, 0.3] (support phase) is allocated, and the weight vector [0.3, 0.7] is allocated in the swing phase. The weight vector is adaptively learned through network training.

[0098] The GRU structure refers to a Gated Recurrent Unit, which is used to capture the time sequence dependence of the posture signal and the plantar pressure signal. The structure is more concise than LSTM, and the training efficiency is higher. For example, the GRU structure contains 32 hidden units, the activation functions of the update gate and the reset gate are sigmoid, and the activation function of the output gate is tanh. The input is the difference feature sequence (dimension 48000) of the posture signal and the plantar pressure signal, and the output is a 64-dimensional motion-force coordination feature vector.

[0099] Different phases of the gait cycle refer to stages divided according to biomechanical characteristics in the gait process, which usually include a support phase (plantar surface in contact with the ground) and a swing phase (plantar surface away from the ground). The physiological signal characteristics of different phases are significantly different. For example, the gait cycle of walking on flat ground is 1.2 seconds, the support phase is 0-0.72 seconds (60%), and the swing phase is 0.72-1.2 seconds (40%). The plantar pressure signal in the support phase has large amplitude and changes dramatically, and the posture signal in the swing phase changes more significantly.

[0100] The adaptive weight vector refers to the weight vector automatically adjusted by the time sequence attention mechanism according to the physiological characteristics of the gait phase, without manual setting, which can better adapt to individual differences. For example, the plantar pressure signal in the support phase of a certain patient has a large deviation, and the adaptive weight vector is [0.8, 0.2], which strengthens the contribution of the plantar pressure signal; another patient's posture signal in the swing phase has a large deviation, and the weight vector is [0.2, 0.8], which strengthens the contribution of the posture signal.

[0101] The clinical rehabilitation evaluation score refers to the score of the commonly used anterior cruciate ligament rehabilitation evaluation scale in clinical practice, which can comprehensively reflect the clinical rehabilitation state of the patient and serve as the basis for weight mapping. For example, the Lysholm knee score scale is used, which includes 8 dimensions such as pain, swelling, lameness, instability, and range of motion, with a full score of 100. A certain patient scored 85, indicating good rehabilitation status.

[0102] The multi-dimensional weight coefficient refers to the weight coefficient corresponding to different evaluation indexes (muscle strength, gait, joint range of motion, fatigue degree) for mapping the clinical rehabilitation evaluation score to, which is used for feature vector fusion. For example, the Lysholm score of 85 points is mapped to the multi-dimensional weight coefficient [0.5 (muscle strength), 0.5 (gait), 0.4 (joint range of motion), 0.6 (fatigue degree)], and the sum of the weight coefficients can not be 1, which is adaptively allocated according to the importance of the evaluation index.

[0103] Specifically: the force-electricity coupling feature gridding mining branch divides the force-electricity coupling difference feature sequence (30 seconds, sampling rate 1600 Hz) into 300 time sequence grid units according to 0.1 seconds, each unit containing 160 data points; the 1D-CNN layer adopts 3 convolution kernels with a size of 3, a step of 1, and a ReLU activation function, extracts the local space-time correlation features (mean, standard deviation, peak value) of each grid unit, and outputs a feature map of 3*298; the bidirectional LSTM layer is set to 64 hidden units, the forward LSTM extracts the time sequence dependence from the 1st unit to the 300th unit, the reverse LSTM extracts the dependence from the 300th unit to the 1st unit, and outputs a 128-dimensional feature, which is averaged to obtain a 64-dimensional gridded force-electricity coupling feature vector; the motion-stress dynamic weight collaborative branch inputs the difference feature sequence of the posture signal and the plantar pressure signal, the time sequence attention mechanism adaptively learns the weight vector (support phase [0.7, 0.3], swing phase [0.3, 0.7]) of different phases of the gait cycle through training, and the GRU structure (32 hidden units) captures the time sequence dependence relationship, the weight vector and the GRU output feature are weighted and fused to obtain a 64-dimensional motion-stress collaborative feature vector; the rehabilitation evaluation score mapping weight layer collects the Lysholm score of 85 points of the patient, maps it to the multi-dimensional weight coefficient [0.5, 0.5, 0.4, 0.6] through the full connection layer, and element-wise weighted fusion is performed on the gridded force-electricity coupling feature vector and the motion-stress collaborative feature vector according to the weight coefficient to generate a 128-dimensional global feature vector, which is used for subsequent exclusive sub-model training.

[0104] The beneficial effects of the above technical solution are: through fine time sequence grid division and double-branch feature extraction, the capture of local space-time features and long-time dependence relationship is strengthened, combined with the adaptive attention mechanism of gait phase, the representation ability of the motion-stress collaborative feature is improved, and through the multi-dimensional weight fusion of the clinical rehabilitation score mapping, the generated global feature vector can more accurately reflect the rehabilitation state of the patient, providing high-quality feature input for subsequent evaluation.

[0105] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these modifications and variations.

Claims

1. A smart wearable rehabilitation monitoring and assessment device for anterior cruciate ligament injury, characterized in that, include: The sensor acquisition module is used to wear different types of sensors on designated locations on the patient and collect data N times at each designated location to obtain N physiological signals for the corresponding designated location. The physiological signals are of four types, including: muscle force signal, surface electromyography signal, posture signal, and plantar pressure signal. The signal processing module is used to filter, enhance, and perform analog-to-digital conversion on each physiological signal to obtain a digital signal, and to construct the signal time series matrix for the corresponding physiological type. The compensation determination module is used to obtain the posture behavior contour set of each patient's current posture, and combine it with the reference behavior contour set of the corresponding standard patient posture to obtain the behavior difference set at the corresponding specified position. The module then performs fitting processing on all behavior difference sets at the same specified position to determine discrete behavior points and the frequency of occurrence of discrete behavior points as compensation factors, and the frequency of occurrence is less than or equal to N. The difference determination module is used to adjust the corresponding row vectors according to the compensation factor and the behavioral contour difference set of each row vector in the signal time series matrix to obtain a first time series matrix, and compare the first time series matrix with the standard time series matrix constructed based on the standard patient posture to obtain a difference matrix, and extract the difference feature vector of the difference matrix; The assessment and analysis module is used to input the differential feature vector of each physiological signal into the muscle strength prediction model, output the quantitative results of the lower limb functional status based on each assessment indicator, and compare all quantitative results with the preset quantitative table to obtain the final assessment result.

2. The intelligent wearable rehabilitation monitoring and assessment device for anterior cruciate ligament injury according to claim 1, characterized in that, The signal processing module includes: The signal interception unit is used to call the start and end interception times of the corresponding physiological signal from the preset mapping library based on the patient's posture at that time, construct a time-domain gate function that is adapted to the feature distribution of the corresponding physiological signal, and match the corresponding physiological signal with the time-domain gate function on the time axis to intercept the target signal segment. The extraction unit is used to extract the in-gate power frequency interference ratio at sampling time t of the target signal segment in real time. Activity segment ratio and periodic stability Calculate the cooperative adaptation feature value at acquisition time t. The calculation formula is: ,in, The preset minimum positive parameter; The normalization determination unit is used to determine the feature distribution dispersion of the target signal segment. The baseline dispersion of the corresponding physiological signal in the preset mapping library is called. The time-domain gate normalization coefficient K is obtained by the following formula: ; The order determination unit is used to determine the order based on the cooperative adaptation feature value. The time-domain gate normalization coefficient K determines the bandpass filter order corresponding to the physiological signal. The calculation formula is: ,in, This is the preset benchmark filter order corresponding to the physiological signal; This is a rounding function; The model adjustment unit is used to match a filtering model from the type-model lookup table that corresponds to the physiological type of the corresponding physiological signal, and based on... The notch frequency of the filtering model is dynamically adjusted, and combined with... Adjust the order of the corresponding filtering model; The filtering unit is used to perform convolution operations between the adjusted model and the time-domain gate function, and to normalize and compensate the corresponding physiological signal in combination with K, and output the corresponding filtered signal.

3. The intelligent wearable rehabilitation monitoring and assessment device for anterior cruciate ligament injury according to claim 2, characterized in that, The signal processing module further includes: The confidence level determination unit is used to extract the signal ratio of the corresponding filtered signal. Core feature retention Combined with the cooperative adaptation feature value of the corresponding filtered signal Determine the dynamic confidence level of the corresponding filtered signal. The calculation formula is: ,in, The baseline co-fit feature value is the standard patient posture matching the corresponding filtered signal in the preset mapping library; The enhancement unit determines the characteristic values ​​of the enhanced signal according to the following formula to obtain the enhanced signal; ,in, Let be the eigenvalue of the enhanced signal at time t; These are the eigenvalues ​​of the filtered signal at time t; The eigenvalue of the baseline feature at time t.

4. The intelligent wearable rehabilitation monitoring and assessment device for anterior cruciate ligament injury according to claim 1, characterized in that, The compensation determination module includes: Set building blocks are used to construct a set of behavioral differences for each patient's current posture at a specified location. ,in, The positional difference of the i1th patient's current posture at the specified location based on the j1th posture point; The first fitting unit is used for each behavior difference set at the same specified location. An internal fitting process is performed to obtain the first linear coefficients and the first discrete points, where the fitting objective function is: ,in, , The slope and intercept of the first linear coefficient; The biomechanical contribution weights are assigned to the j1th attitude point; The first statistical unit is used to count the first occurrence frequency of the first discrete point of the same pose point at the same specified location based on all behavior difference sets; The second fitting unit is used to perform linear fitting analysis on all first occurrence frequencies to obtain second linear coefficients and second discrete points, and to count the second occurrence frequency of the second discrete points of the same pose point at the same specified position based on all behavior difference sets. The threshold determination unit is used to construct a frequency threshold for the j1th attitude point based on the first occurrence frequency, the second occurrence frequency, the first linear coefficient, and the second linear coefficient. ; The factor determination unit is used to filter out third occurrence frequencies that are greater than a corresponding frequency threshold from all first occurrence frequencies, and to treat attitude points that match the third occurrence frequencies as discrete behavior points to obtain compensation factors. ,in, This represents the third occurrence frequency of the j1th attitude point.

5. The intelligent wearable rehabilitation monitoring and assessment device for anterior cruciate ligament injury according to claim 1, characterized in that, The threshold determination unit includes: ； in, The biomechanical weighted average of the first occurrences is given, and , The frequency of the first occurrence of the j1th attitude point; The biomechanical weighted mean of the second occurrence frequency, and , The second occurrence frequency of the j1th attitude point; This corresponds to the coefficient of the patient's current recovery stage; The confidence level for fitting the objective function; The slope based on the second linear coefficient The coefficient of determination of the intercept.

6. The intelligent wearable rehabilitation monitoring and assessment device for anterior cruciate ligament injury according to claim 1, characterized in that, The difference determination module includes: The difference determination unit is used to obtain the timestamp and attitude identifier corresponding to each row vector in the signal time sequence matrix, and perform dual matching of the behavior contour difference set at the same specified position in terms of time axis and attitude dimension to determine the target behavior contour difference set corresponding to each row vector. The element adjustment unit is used to extract the effective discrete behavioral points and real-time weighted frequencies in the compensation factor, and combine them with the corresponding biomechanical phase weights of the anterior cruciate ligament rehabilitation biomechanical phase of the corresponding behavioral points to determine the feature difference degree between each row vector and the corresponding target behavioral contour difference set, and obtain the adjustment intensity coefficient for adjustment. The recombination unit is used to recombine the adjusted row vectors according to the original temporal matrix arrangement to form the first temporal matrix.

7. The intelligent wearable rehabilitation monitoring and assessment device for anterior cruciate ligament injury according to claim 1, characterized in that, Also includes: The vector generation module is used to generate a global feature vector by mining branches based on the gridded feature of force-electric coupling, coordinating branches of motion-force dynamic weights, and setting a weight layer for the rehabilitation assessment score mapping at the output of the dual branches. The sub-model training module is used to train and construct a dedicated sub-model that corresponds one-to-one with each lower limb function assessment index based on global feature vectors, serving as a muscle strength prediction model.

8. The intelligent wearable rehabilitation monitoring and assessment device for anterior cruciate ligament injury according to claim 1, characterized in that, The vector generation module includes: The force-electric coupling feature gridding mining branch is used to divide the force-electric coupling difference feature sequence into fixed-length temporal grid units. It uses 1D-CNN and bidirectional LSTM structure to output gridded force-electric coupling feature vectors. The 1D-CNN layer extracts the local spatiotemporal correlation features within each grid unit, and the bidirectional LSTM layer captures the long-term dependencies between grid units. The motion-force dynamic weighted coordination branch is used to output motion-force coordinated feature vectors using a temporal attention mechanism with dynamic weight vectors and a GRU structure. The temporal attention mechanism assigns adaptive weight vectors based on physiological features to the difference features of different phases within the gait cycle. The weight vectors are weighted and fused with the output features of the GRU layer to model the dynamic coordination relationship between posture and plantar pressure signals. The rehabilitation assessment score mapping weight layer is used to collect patients' clinical rehabilitation assessment scores, map the scores to multi-dimensional weight coefficients, and assign fusion weights to the dual-branch feature vectors to generate a global feature vector.

Images