A knee joint dysfunction classification and identification method and system based on knee joint gait kinematics

Abstract

The application belongs to the technical field of gait analysis, and is a knee joint dysfunction classification and recognition method and system based on knee joint gait kinematics, which comprises the following steps: collecting multi-degree-of-freedom kinematic time series signals of bilateral knee joints; pre-processing the time series signals, and screening the time series signals according to rotation and translation degrees of freedom; performing quality control on the screened time series signals based on gait cycle integrity and repeated cycle consistency, and identifying incomplete cycles or abnormal cycles; performing data enhancement processing on the time series signals to expand the training samples; constructing a multi-channel feature fusion classification network model, training the model using the expanded training sample set, classifying the time series signals using the trained model, and obtaining the classification prediction result of knee joint function. The application combines gait field knowledge with collection variability modeling, and realizes the whole process processing from gait kinematic signal collection to knee joint dysfunction classification.

Description

Technical Field

[0001] This invention belongs to the field of gait analysis technology, specifically relating to a method and system for classifying and identifying knee joint dysfunction based on knee joint gait kinematics. Background Technology

[0002] Knee dysfunction is a leading cause of chronic pain, limited mobility, and disability. Its identification and classification workflow typically relies on structured assessments tailored to specific functional impairments, with key technical approaches including: imaging techniques (MRI, X-rays, etc.), traditional machine learning techniques based on handcrafted features, and general-purpose temporal deep learning techniques. These existing technical solutions suffer from the following drawbacks: Structured imaging technology relies on expensive hardware infrastructure and highly specialized human resources, making large-scale pre-deployment difficult. In resource-scarce areas, equipment availability and expert interpretation are easily limited or delayed. Different studies have varying acquisition protocols, feature definitions, and classification goals, lacking large-scale, multi-center unified datasets and discriminative frameworks for multiple types of functional impairments. Traditional machine learning methods rely on manual feature extraction, but manual feature engineering loses the dynamic information of the original temporal sequence during dimensionality reduction, making it difficult to characterize continuous spatiotemporal structures. General temporal deep learning models lack gait domain priors, do not inject gait-specific biomechanical structural inductive biases, and do not explicitly model multi-center acquisition variability, resulting in poor cross-center generalization ability. Existing deep models do not perform symmetrical modeling of the bilateral anatomical structures of the lower limbs, cannot distinguish between unilateral local abnormalities and bilateral compensatory asymmetry, and do not inject explicit biomechanical knowledge into the feature extraction process, lacking biomechanical interpretability. Summary of the Invention

[0003] To address the problems existing in the prior art, this invention provides a method and system for classifying and identifying knee joint dysfunction based on knee joint gait kinematics. By combining gait domain knowledge, including multi-degree-of-freedom structures, limb compensation relationships, gait phase organization, frequency domain rhythms, and explicit knee joint motion function descriptors, with data acquisition variability modeling, the entire process from gait kinematics signal acquisition to knee joint dysfunction classification is realized.

[0004] On one hand, embodiments of the present invention provide a method for classifying and identifying knee joint dysfunction based on knee joint gait kinematics, comprising the following steps:

[0005] S1. Acquire multi-degree-of-freedom kinematic time-series signals of bilateral knee joints. The kinematics of each knee joint is described by a multi-dimensional characterization that includes at least rotational and translational degrees of freedom.

[0006] S2. Standardize and preprocess the collected bilateral knee joint multi-degree-of-freedom kinematic time-series signals, and screen the kinematic range of the time-series signals according to the rotational and translational degrees of freedom to identify physically unreliable kinematic trajectories and remove bilateral knee joint multi-degree-of-freedom kinematic time-series signal samples with abnormal rotational or translational degrees of freedom.

[0007] S3. Perform quality control on the time-series signals after kinematic range screening based on gait cycle integrity and repetition cycle consistency to identify incomplete cycles or morphologically abnormal cycles.

[0008] S4. Perform data augmentation on the preprocessed and quality-controlled time-series signal to obtain augmented data, which is used to expand the training samples.

[0009] The data augmentation process involves simulating different types of reasonable acquisition perturbations based on the acquisition method and physical structure of the multi-degree-of-freedom kinematic time-series signal of the knee joint, and expanding the kinematic performance forms that may correspond to the same training sample while keeping the original training sample category labels unchanged.

[0010] S5. Construct a multi-path feature fusion classification network model, and train the constructed classification network model using an expanded training sample set; use the trained classification network model to classify time-series signals to obtain the classification prediction results of knee joint function.

[0011] On the other hand, embodiments of the present invention also provide a knee joint dysfunction classification and recognition system based on knee joint gait kinematics, implemented using the above-mentioned knee joint dysfunction classification and recognition method, including:

[0012] The data acquisition module collects multi-degree-of-freedom kinematic time-series signals of both knee joints. The kinematics of each knee joint is described by a multi-dimensional characterization that includes at least rotational and translational degrees of freedom.

[0013] The preprocessing module performs standardized preprocessing on the collected bilateral knee joint multi-degree-of-freedom kinematic time-series signals, and screens the kinematic range of the time-series signals according to the rotational and translational degrees of freedom to identify physically unreliable kinematic trajectories and remove bilateral knee joint multi-degree-of-freedom kinematic time-series signal samples with abnormal rotational or translational degrees of freedom.

[0014] The quality control module performs quality control on the time-series signals after kinematic range screening based on gait cycle integrity and repetition cycle consistency, and identifies incomplete cycles or morphologically abnormal cycles.

[0015] The data augmentation module performs data augmentation on the preprocessed and quality-controlled time-series signals to obtain augmented data, which is used to expand the training samples.

[0016] The data augmentation process involves simulating different types of reasonable acquisition perturbations based on the acquisition method and physical structure of the multi-degree-of-freedom kinematic time-series signal of the knee joint, and expanding the kinematic performance forms that may correspond to the same training sample while keeping the original training sample category labels unchanged.

[0017] A multi-path feature fusion classification network model is trained using an expanded training sample set. The trained classification network model is then used to classify time-series signals to obtain classification prediction results for knee joint function.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] 1. This invention uses a bi-branch structure that extracts unilateral local anomalies and models bilateral compensation relationships to capture motor dysfunction caused by two mechanisms: unilateral local anomalies and bilateral compensation asymmetry. It designs a classification network that integrates three complementary representations: time domain, frequency domain, and explicit kinematics, and unifies data-driven implicit representations with explicit representations based on biomechanical knowledge. Through preprocessing and quality control processes, it ensures the physical rationality of gait signals and the integrity of gait cycles.

[0020] Overall, a non-invasive, scalable, and low-cost classification and recognition system has been constructed. It uses gait kinematics as its data foundation, does not require expensive imaging equipment such as MRI / X-ray, and can be deployed on a large scale in resource-constrained scenarios such as primary care, communities, and sports fields, serving as an objective functional screening layer before imaging diagnosis.

[0021] 2. This invention establishes a unified discrimination framework based on bilateral multi-degree-of-freedom knee joint gait kinematics to distinguish between healthy knee joint function, injury-related knee joint dysfunction, and degenerative knee joint dysfunction. For the first time, it incorporates two different types of knee joint dysfunction—injury-related and degenerative—into the same discrimination framework, overcoming the limitations of the existing binary classification paradigm of single functional impairment and healthy knee joint dysfunction.

[0022] 3. This invention introduces a data acquisition and perception enhancement strategy to simulate the variability of multi-center acquisition. By injecting gait domain knowledge such as multi-degree-of-freedom structure, phase organization, frequency domain rhythm, and explicit descriptors, and enhancing acquisition and perception, the model's F1 score on the independent external test set only decreased from 0.9236 to 0.9156 on the internal test set, while the baseline TimesNet dropped sharply from 0.7672 to 0.6098 and PatchTST from 0.8401 to 0.7695, thus improving the model's cross-center generalization ability.

[0023] 4. This invention has biomechanical interpretability: By explicitly encoding six knee joint function descriptors through KFE, the model decision-making is aligned with the experience of knee joint dysfunction assessment; through TFE attention visualization, the gait phase that the model focuses on (such as the weight-bearing reaction period and the terminal swing period) can be located, and it was found that the model attention of people with injury-related knee joint dysfunction is significantly focused on the dysfunction side, so as to provide interpretable model features. Attached Figure Description

[0024] Figure 1 This is a flowchart of the overall process for classifying and identifying knee joint dysfunction in an embodiment of the present invention;

[0025] Figure 2 This is the overall diagram of the PACENet three-way fusion architecture;

[0026] Figure 3 This is a diagram of the UFE / CFE dual-branch structure of TFE. Detailed Implementation

[0027] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments, but the implementation of the present invention is not limited thereto.

[0028] Example

[0029] Reference Figures 1-3 This embodiment provides a method for classifying and identifying knee joint dysfunction based on knee joint gait kinematics, including the following steps:

[0030] Step S1: Acquire multi-degree-of-freedom kinematic timing signals of both knee joints.

[0031] Multi-degree-of-freedom kinematic temporal signals of the left and right knee joints were collected during the subjects' walking process. The kinematics of each knee joint was described by a multi-dimensional representation including at least rotational and translational degrees of freedom.

[0032] In this embodiment, each knee joint is characterized by six degrees of freedom (6-DOF), including three rotational degrees of freedom and three translational degrees of freedom. The rotational degrees of freedom include flexion / extension (sagittal plane), adduction / adduction (coronal plane), and internal / external rotation (transverse plane). The translational degrees of freedom include anterior-posterior translation, medial-lateral translation, and proximal-distal translation.

[0033] Therefore, the bilateral knee joint multi-degree-of-freedom kinematic timing signals acquired in this embodiment are 12-channel timing signals, denoted as:

[0034] ;

[0035] in This indicates the number of time steps. The first 6 channels correspond to the kinematic timing signal of the left knee joint, and the last 6 channels correspond to the kinematic timing signal of the right knee joint.

[0036] ;

[0037] Step S2: Standardize and preprocess the collected bilateral knee joint multi-degree-of-freedom kinematic time-series signals. Screen the kinematic range of the time-series signals according to the rotational and translational degrees of freedom to identify physically unreliable kinematic trajectories and remove bilateral knee joint multi-degree-of-freedom kinematic time-series signal samples with abnormal rotational or translational degrees of freedom.

[0038] Specifically, it includes the following sub-steps:

[0039] S21, Two-stage zero-phase filter.

[0040] To reduce acquisition-related noise and preserve the timing structure of the gait waveform, two stages of filtering are applied sequentially to the timing signal:

[0041] The first stage is a Kaiser window FIR band-stop filter used to suppress narrowband interference. Let the sampling frequency be... Nyquist frequency Given center frequency With bandwidth The stopband boundary is defined as:

[0042] ;

[0043] The ideal band-stop frequency response is:

[0044] ;

[0045] A finite-length FIR approximation is obtained by adding windows to the Kaiser window:

[0046] ;

[0047] The second stage is a Butterworth low-pass filter, used to attenuate high-frequency sensor jitter and motion artifacts. (Regarding the cutoff frequency...) order The Butterworth filter has the following amplitude response:

[0048] ;

[0049] Both stages of filtering employ zero-phase forward and reverse filtering to avoid phase distortion. The equivalent frequency response after zero-phase filtering is: The complete filtering process can be represented as:

[0050] ;

[0051] in This indicates a zero-phase FIR band-stop filter. This indicates a zero-phase Butterworth low-pass filter.

[0052] S22, Screening range of kinematics.

[0053] After filtering in step S21, the kinematic range of the six-degree-of-freedom kinematic time-series signals of the left and right knee joints of each subject is further screened to identify physically unreliable kinematic trajectories caused by factors such as tracking marker occlusion or displacement, loose rigid body fixation, and anatomical calibration deviation. For the first... The first subject, whose first The kinematic timing signal of the lateral knee joint is represented as follows: ,in , Indicates the left knee joint. This represents the right knee joint. 600 represents the number of time sampling points retained after preprocessing, and 6 represents the six degrees of freedom of the knee joint. The six degrees of freedom are adduction and abduction, internal and external rotation, flexion and extension, anterior and posterior translation, vertical translation, and lateral translation.

[0054] More preferably, before determining the kinematic range, the data structure of the kinematic data for each knee joint is first verified. If the number of time points or degrees of freedom of the kinematic data for a certain knee joint does not meet the predetermined requirements... If the structure is abnormal, it indicates that the kinematic data of that knee joint may be incomplete, have missing channels, be misaligned, or have failed to be parsed. The kinematic data of that knee joint will be marked as structurally abnormal. This structural verification helps prevent data with incorrect channel order or inconsistent lengths from being mistakenly used as normal signals in subsequent calculations.

[0055] After structural verification, the six-degree-of-freedom kinematic time-series signal was divided into rotational and translational degrees of freedom according to physical properties, and different screening methods were used for each. The first three columns are rotational degrees of freedom, corresponding to adduction / abduction, internal / external rotation, and flexion / extension, respectively; the last three columns are translational degrees of freedom, corresponding to forward / backward translation, vertical translation, and inward / outward translation, respectively.

[0056] S221. Rotational degree of freedom anomaly screening: Statistically analyze the absolute values ​​of the angles of the rotational degrees of freedom of the time sequence signal throughout the entire acquisition time sequence, and calculate the maximum absolute angle; compare the calculated maximum absolute angle with the preset upper limit threshold of the corresponding motion degree of freedom to determine whether the rotational degree of freedom of the knee joint on the corresponding side of the time sequence signal is abnormal.

[0057] For the The first subject The side of the first There are 1 rotational degree of freedom, and their timing signals are denoted as . Considering that the rotation signal may exhibit angular changes in both positive and negative directions simultaneously, calculating only the maximum or minimum value might miss abnormal angles in the opposite direction. Therefore, the absolute values ​​of the rotational degrees of freedom of the kinematic time-series signal are statistically analyzed throughout the entire acquisition time series, and the maximum absolute angle is calculated:

[0058] ;

[0059] in, Indicates the time sampling point. These correspond to the three rotational degrees of freedom: adduction / abduction angle, internal / external rotation angle, and flexion / extension angle. The calculated maximum absolute angle... Compare with the preset upper limit threshold for the corresponding degree of freedom of motion. The preset upper limit threshold for rotation is:

[0060] ;

[0061] Among them, the three thresholds of the upper limit of rotation These correspond sequentially to the adduction / abduction angle, internal / external rotation angle, and flexion / extension angle. If the following conditions are met... This indicates that the corresponding rotational degree of freedom has an angle value exceeding the preset physical range at at least one time point, and the knee joint on the side corresponding to the timing signal is judged to have an abnormal rotational degree of freedom. In other words, as long as the maximum absolute value of the adduction and abduction angle exceeds... The maximum absolute value of the internal and external rotation angles exceeds Or the maximum absolute value of the flexion / extension angle exceeds This is marked as an anomaly in rotational degrees of freedom.

[0062] S222, Translational Degree of Freedom Anomaly Screening: Calculate the difference between the maximum and minimum values ​​of the translational degree of freedom of the time-series signal throughout the entire acquisition time sequence to obtain the translational motion range; compare the calculated translational motion range with the preset translational range threshold of the corresponding motion degree of freedom to determine whether the knee joint translational degree of freedom on the corresponding side of the time-series signal is abnormal.

[0063] For the The first subject The side of the first There are translational degrees of freedom, and their time-series signals are denoted as . Because the absolute position of the translation signal may be affected by the initial coordinate origin, calibration reference, and installation position, directly using the maximum absolute value of the translation signal may misjudge the overall baseline offset as abnormal. Therefore, this step does not use the absolute position at a single moment as the screening criterion, but instead calculates the difference between the maximum and minimum values ​​of the translational degrees of freedom of the time-series signal throughout the entire acquisition time series to obtain the translational motion range:

[0064] ;

[0065] in, Indicates the time sampling point. These correspond to three degrees of freedom in translation: forward / backward, vertical / vertical, and inward / outward. The calculated range of translational motion... It is compared with the preset translation range threshold for the corresponding degree of freedom of motion. If the translation range is greater than the preset translation range threshold, that is... If so, the knee joint translational degree of freedom on the side corresponding to the timing signal is determined to be abnormal. In this embodiment, the translational range threshold is preset to:

[0066] ;

[0067] The three thresholds correspond to forward / backward translation, vertical translation, and inward / outward translation, respectively. If the following conditions are met... If the total displacement range of the translational degree of freedom exceeds the preset range during the acquisition period, the corresponding knee joint sample will be judged as having an abnormal translation range. In other words, if the difference between the maximum and minimum translation values ​​in any of the forward / backward, up / down, or inward / outward directions exceeds the threshold of 8 cm, it will be marked as a translational anomaly.

[0068] S223. Completely remove the bilateral knee joint multi-degree-of-freedom kinematic timing signal samples of the subjects.

[0069] To ensure structural symmetry in subsequent bilateral knee joint multi-degree-of-freedom kinematic modeling, if either knee joint triggers any of the following rejection criteria, the entire bilateral knee joint multi-degree-of-freedom kinematic temporal signal samples of the subject will be rejected:

[0070] 1. Either the left or right knee joint's kinematic matrix does not conform to the predetermined... Matrix structure;

[0071] 2. The maximum absolute angle of any rotational degree of freedom of the left or right knee joint exceeds the corresponding upper limit threshold for rotation;

[0072] 3. The range of motion of any translational degree of freedom of the left or right knee joint exceeds the corresponding translational range threshold.

[0073] The above screening was performed on the multi-degree-of-freedom kinematic temporal signals of the left and right knee joints respectively, but the final exclusion operation was performed at the subject level. If the... If a subject's left or right knee joint triggers any of the above exclusion criteria, the subject is marked as unqualified for kinematic range screening, and the multi-degree-of-freedom kinematic signal samples of the left and right knee joints are excluded from the subsequent processing as a whole.

[0074] S3. Adaptive Gait Cycle Segmentation and Morphological Similarity Quality Control: Performs quality control on time-series signals after kinematic range screening based on gait cycle integrity and repetition cycle consistency to identify incomplete cycles or morphologically abnormal cycles.

[0075] After the kinematic range screening in step S22, the kinematic time-series signals of the left and right knee joints of each subject are further subjected to quality control based on the integrity of gait cycles and the consistency of repetition cycles. This identifies incomplete or abnormally shaped cycles caused by local signal loss, tracking errors, abnormal body swings, or brief interference during the acquisition process, in order to solve the problem of inconsistent walking speed, gait rhythm, and number of effective gait cycles among different subjects.

[0076] Unlike the traditional method of extracting kinematic signals according to a fixed time length, this step first estimates the gait cycle based on the periodic characteristics of the flexion and extension motion signals (i.e., flexion and extension angle signals) of each knee joint. Then, the cycle boundary is determined based on the estimation results of the gait cycle. Representative complete cycles are selected through cycle time normalization and morphological similarity evaluation to achieve cycle segmentation. After performing gait cycle estimation and cycle segmentation on the six-degree-of-freedom kinematic time-series signals of the left and right knee joints respectively, the structural pairing of the bilateral knee joint cycles is completed based on the number of effective cycles of both knee joints.

[0077] For the The subjects' bilateral knee joint multi-degree-of-freedom kinematic time-series signals were represented as follows: ,in This represents the number of time sampling points; the first six channels are the six-degree-of-freedom kinematic timing signals of the left knee joint, and the last six channels are the six-degree-of-freedom kinematic timing signals of the right knee joint, i.e. , The six degrees of freedom of each knee joint correspond to the adduction and abduction angles, internal and external rotation angles, flexion and extension angles, anterior-posterior translation, superior-inferior translation, and lateral-inferior translation, respectively. Since the knee flexion and extension angle signals usually exhibit relatively obvious periodic changes during continuous walking, the flexion and extension angle signals of each knee joint are selected as reference signals for gait period estimation and period boundary detection.

[0078] Specifically, step S3 includes:

[0079] S31. Utilizing the repetitive variation pattern of the complete flexion-extension angle signal, and based on the autocorrelation function of the flexion-extension angle signal and the number of oscillation peaks in the flexion-extension angle signal, the gait cycle is adaptively estimated.

[0080] For the For each subject, the flexion-extension angle signal of either knee joint is recorded as:

[0081] ;

[0082] Since the flexion-extension angle may have an overall bias caused by the initial standing posture, anatomical calibration, or measurement reference, the average value of the flexion-extension angle signal at all time sampling points is first calculated, and then this average value is subtracted from the original flexion-extension angle signal to obtain the mean-reduced centered signal:

[0083] ;

[0084] Mean removal can reduce the influence of the signal baseline position on periodic evaluation, so that subsequent autocorrelation calculations mainly reflect the time structure of the flexion-extension angle signal as the gait repeats, rather than the absolute angle level of the signal.

[0085] Subsequently, the autocorrelation function under different candidate time delays is calculated for the centered flexure / extension angle signal:

[0086] ;

[0087] in, This represents the candidate time delay, and also the number of sampling points corresponding to the candidate gait cycle. To avoid incorrectly identifying cycles in excessively short local fluctuations or excessively long slow trends, the candidate time delay is limited to a preset range of physiologically reasonable gait cycles.

[0088] ;

[0089] in, This represents the sampling frequency of the kinematic signal. Within the above candidate delay range, the delay corresponding to the maximum value of the autocorrelation function is taken as the period estimation result based on the overall periodicity repetition:

[0090] .

[0091] Autocorrelation estimation uses the repetitive variation pattern of the complete flexion-extension angle signal to determine the period. When multiple complete gait cycles exist in the flexion-extension angle signal, similar waveforms in adjacent gait cycles will generate large autocorrelation values ​​at the corresponding time delays. To reduce the potential impact of waveform harmonics or local abnormal fluctuations on relying solely on the autocorrelation function, this step also obtains another period estimation result based on the number of oscillation peaks in the flexion-extension angle signal.

[0092] First, calculate the dynamic range of the flexion / extension angle signal:

[0093] ;

[0094] Then, a peak significance threshold is set based on the dynamic range of the flexion-extension angle signal, and candidate oscillation peaks are detected. The significance condition used for the initial peak detection is:

[0095] ;

[0096] The constant 5 is used to set the minimum peak significance. This is used to adapt the peak detection conditions to different subjects' knee flexion and extension ranges. Taking the larger of the two values ​​avoids misidentifying minute local fluctuations as valid swing peaks when the knee flexion and extension range is small or noise is significant. The set of detected swing peak positions is further represented as:

[0097] ;

[0098] in, For the first The position of the swing peak The number of swing peaks detected.

[0099] When at least two swing peaks are detected, i.e. At that time, the period estimation result based on peak count is calculated according to the length of the complete flexion-extension angle signal and the number of oscillation peaks:

[0100] ;

[0101] In the formula This refers to periodic estimation based on peak counts, or simply peak count estimation.

[0102] When at least two valid oscillation peaks are detected, the period estimate obtained through the autocorrelation function and the period estimate obtained based on the number of oscillation peaks are combined, and the minimum value of the two period estimates is taken as the final adaptive gait period estimate, which can be expressed as:

[0103] ;

[0104] When fewer than two oscillation peaks are detected, the peak count is insufficient to support period estimation, and only autocorrelation estimation is used in this case:

[0105] .

[0106] S32. Based on the gait cycle estimate obtained in step S31, set the peak significance condition and the minimum distance between adjacent peaks, and perform peak detection on the flexion-extension angle signal to determine the cycle boundary.

[0107] Obtain adaptive gait period estimate Then, based on the adaptive gait period estimate A second peak detection is performed on the flexion-extension angle signal. Unlike the initial peak detection in step S231, the second peak detection not only sets a peak significance condition but also a minimum distance between adjacent peaks to reduce the problem of multiple local peaks being repeatedly identified within the same oscillation phase. The second peak detection satisfies:

[0108] ;

[0109] in, The minimum distance between adjacent peaks, i.e., the minimum peak spacing, is adaptively determined based on the current gait cycle estimate of the lateral knee joint; constant 3 and Together, they constitute the peak salience threshold for the second peak detection. Compared to the first peak detection, the second peak detection is used to determine the periodic boundary under the condition of a known gait periodic scale. Therefore, a relatively low salience condition is adopted under the inter-peak distance constraint to balance the stability and completeness of peak recognition.

[0110] For the two adjacent swing peaks obtained from the second peak detection and A local search interval is set between the two, and the local minimum value of the flexion-extension angle signal is found within this local search interval as the periodic tangent point. The starting position of the local search interval is set as follows:

[0111] ;

[0112] The above starting position setting places the search starting point after the oscillation peak, avoiding mistaking short-term fluctuations at or near the peak of the oscillation peak as period boundaries. This applies to adjacent oscillation peaks. and Further detection of local peaks; relative to the detected local peaks The time offset is denoted as Then the termination position of the local search interval is set as follows:

[0113] ;

[0114] In the local search interval Within the range, find the time position where the flexion-extension angle signal value is minimum, and determine it as the first time position. Each cycle cut-off point:

[0115] ;

[0116] Using local minima instead of directly using the oscillation peak as the periodic cut-off point allows adjacent periods to be segmented at relatively stable valley positions in the flexure / extension angle signal, reducing the impact of oscillation peak height variations or irregular peak shapes on the start and end positions of the period. This embodiment cuts off two adjacent periods at... and The data between these points is defined as a primitive gait cycle. For the six-DOF motion timing signal of either knee joint... , No. The original cycle is represented as:

[0117] ;

[0118] The lengths of the original cycles obtained in this way can be different, and their lengths are denoted as follows: ,Right now: .

[0119] S33. Perform linear interpolation on the estimated gait period to normalize the gait period time.

[0120] Since the duration of gait cycles may vary among different subjects and even among different subjects within the same subject, directly comparing raw cycles of different lengths can lead to misalignment of corresponding gait phases. Therefore, each raw estimated gait cycle is linearly interpolated along the time dimension to unify it into a normalized cycle length. For the original estimated gait period Linear interpolation was performed on each of the six degrees of freedom channels to make the first and last time points in the original estimated gait cycle correspond to the start and end points of the normalized cycle, respectively. The remaining time points were mapped according to their relative cycle positions to obtain the time-normalized gait cycle. .

[0121] S34. Based on the gait cycles after time normalization of the left and right knee joints, count the number of effective normalized cycles extracted from the left and right knee joints, and take the smaller value as the number of normalized paired cycles for bilateral knee joint modeling.

[0122] The aforementioned period estimation, boundary detection, and time normalization processes were applied to the left and right knee joints, respectively. (Statistical analysis follows.) Number of effective normalized cycles extracted from the left knee joint of each subject And the number of effective normalized cycles extracted from the right knee joint. The smaller of the two values ​​is taken as the normalized number of paired cycles for bilateral knee joint modeling:

[0123] ;

[0124] For the For each paired period, the normalized periods of the left and right knee joints are concatenated along the channel dimension according to the order in which the periods of the left and right knee joints appear, to obtain the normalized paired periods of both knee joints:

[0125] ;

[0126] in This represents the normalized periodic sequence of the left knee joint. This represents the normalized periodic sequence of the right knee joint. If no valid complete period is extracted from either the left or right knee joint, this indicates the number of paired periods in the bilateral knee joint modeling. Therefore, it is impossible to construct a normalized paired cycle for both knee joints. At this point, the first... One subject was deemed to have an incomplete cycle, and their bilateral data were removed from the subsequent processing.

[0127] S35. Evaluate the morphological similarity between normalized paired cycles of different knee joints of the same subject; based on morphological similarity, screen abnormal cycles to obtain the actual number of complete cycles of both knee joints retained by the subject; if the actual number of complete cycles of both knee joints retained is greater than the preset minimum number of effective cycles, the subject is deemed to have passed the quality control of gait cycle integrity and morphological consistency.

[0128] Even if the cycle boundaries can be successfully detected, some cycles may still be affected by brief tracking loss, abnormal gait, body posture adjustments, or local signal jumps. Such cycles may be complete in duration, but their multi-degree-of-freedom motion patterns may deviate significantly from the repetitive gait patterns of the same subject. Therefore, this embodiment further evaluates the morphological similarity between normalized paired cycles of different knee joints of the same subject after gait cycle time normalization.

[0129] For the Extracted from each subject Bilateral knee joint normalized pairing cycle , each The periodic matrix is ​​expanded into a one-dimensional paired periodic vector according to a predetermined channel order and time order:

[0130] ;

[0131] Calculate the Pearson correlation coefficient between any two paired period vectors of the same subject:

[0132] ;

[0133] in, Indicates the first The pairing periodic vector and the first The degree of morphological correlation between the paired periodic vectors. For the ... The nth paired cycle vector is used to calculate the average of its correlation coefficients with all paired cycles of the corresponding subject, thus obtaining the nth pairing cycle vector. Average morphological similarity of paired periodic vectors:

[0134] ;

[0135] Cycles with high average morphological similarity indicate that they maintain a relatively consistent bilateral movement pattern with most cycles of the same subject; cycles with low average morphological similarity are considered to have a stronger randomness or abnormality.

[0136] Normalized number of pairing periods Since the number of cycles available for comparison is limited, directly deleting a cycle may cause the subject to lose enough complete cycles, so all of their paired cycles are retained.

[0137] when At that time, according to the preset discard ratio Calculate the number of cycles to be discarded:

[0138] ;

[0139] in Indicates the percentage of items discarded.

[0140] The corresponding number of retention periods is:

[0141] ;

[0142] in This indicates the number of cycles with the lowest similarity that were discarded.

[0143] Based on average morphological similarity Sort all pairing cycles from highest to lowest and select the ones with the highest average morphological similarity. These cycles constitute the set of retention cycles:

[0144] ;

[0145] After completing the morphological similarity screening, the statistics are as follows: The actual number of complete bilateral knee joint cycles retained by each subject and compare it with the preset minimum number of effective cycles. If a comparison is made, If the gait cycle completeness and morphological consistency are satisfactory, the subject is deemed to have passed quality control, and the retained cycles are used for subsequent kinematic feature extraction and knee joint function classification. Otherwise, it indicates that after removing incomplete or morphologically abnormal cycles, the number of remaining effective cycles is insufficient to stably represent the subject's repetitive gait pattern. In this case, the subject is deemed to have unsatisfactory cycle quality, and all of the subject's left and right knee joint multi-degree-of-freedom kinematic time-series signal samples are removed from the subsequent processing.

[0146] Step S4: Perform data augmentation on the kinematic time-series signal after preprocessing and quality control to obtain augmented data, which is used to expand the training samples of the model.

[0147] In actual gait data acquisition, even if different subjects complete the same walking test, their kinematic signals may still vary due to walking speed, range of motion, limb compensation methods, and individual anatomical differences. Furthermore, different acquisition centers or batches may exhibit non-pathological signal differences due to variations in rigid body or marker placement, anatomical calibration results, sensor coordinate system orientation, equipment background noise, and initial zero-point deviation. To improve the model's adaptability to data differences between different subjects, batches, and acquisition centers, acquisition-sensory data augmentation is performed on the preprocessed and quality-controlled kinematic temporal signals during the model training phase.

[0148] If the model learns using only the original training samples, it may mistakenly identify the aforementioned acquisition conditions or individual differences as disease-related features, thereby reducing the model's recognition stability on new subjects or at new acquisition centers. Therefore, this step does not involve unconstrained transformation of the kinematic data, but rather simulates different types of reasonable acquisition perturbations based on the acquisition method and physical structure of the knee joint's six-degree-of-freedom kinematic time-series signals. While keeping the original training sample category labels unchanged, this expands the possible kinematic representations corresponding to the same training sample.

[0149] Let the original multi-channel kinematic time-series signal of a training sample be... ,in For the number of channels, The number of time sampling points; the first The timing signals of each channel are denoted as follows: .

[0150] This step sets up six enhancement operations: crosstalk disturbance, global scaling, amplitude distortion, time distortion, random bias, and jitter. Different enhancement operations change the channel coupling relationship, overall amplitude, local amplitude, time progression, baseline position, and local random fluctuations of the kinematic time series signal, respectively, to correspond to different sources of actual acquisition variation.

[0151] The data before and after augmentation have the same number of channels and duration, and inherit the subject class labels from the original training samples. The augmented data is used only for model parameter training and is not used for model performance evaluation on the validation and test sets.

[0152] S41. Add crosstalk disturbance.

[0153] Crosstalk disturbances are used to simulate slight alignment errors between the sensor coordinate system and the anatomical coordinate system caused by deviations in the installation position of rigid bodies or markers.

[0154] In actual data acquisition, knee joint rotational motion needs to be reconstructed based on rigid bodies or markers fixed to the femur and tibia. When there are slight deviations in the rigid body mounting direction, marker positions, or anatomical key point calibration results, the coordinate axes of the device coordinate system may not completely coincide with the predetermined anatomical coordinate axes. In this case, the rotational motion that actually occurs in a certain anatomical motion plane may be partially decomposed into two other rotational channels, manifesting as signal leakage or crosstalk between rotational degrees of freedom.

[0155] For example, changes in angle that are primarily manifested as flexion and extension movements may, with a slight shift in the coordinate system, simultaneously generate small additional components in the adduction / abduction or internal / external rotation channels. This change does not necessarily indicate that the subject has actually performed additional abnormal movements, but may originate from a slight alignment difference in the acquisition coordinate system.

[0156] To simulate the above phenomenon, three rotational degrees of freedom were extracted from the left and right knee joints respectively. For the first... The three-dimensional trajectory of the lateral knee joint is represented as follows:

[0157] ;

[0158] The three channels correspond to the adduction / abduction angle, internal / external rotation angle, and flexion / extension angle, respectively. Small Euler angles about the three coordinate axes are sampled respectively. , and And construct the rotation matrix:

[0159] ;

[0160] Apply the rotation matrix to the three-dimensional angular trajectory of the corresponding knee joint:

[0161] ;

[0162] This process is equivalent to applying a small coordinate system rotation to the original three-dimensional angle signal, so that the three angle channels are recombined according to the proportions determined by the rotation matrix.

[0163] Before adding crosstalk disturbance, the three rotating channels are represented according to the original acquisition coordinate system; after adding crosstalk disturbance, the signal duration, gait rhythm and overall motion trend remain unchanged, but the components corresponding to small-angle coordinate deflection are redistributed between different rotating channels.

[0164] This step only performs a joint transformation on the three rotational degrees of freedom on the same side, without mixing the translational degrees of freedom into the angle subspace, and without performing unrelated random additions or subtractions on the three angle channels separately. This preserves the geometric correlation between the three-dimensional rotational components, making the enhancement result closer to the coordinate crosstalk caused by actual rigid body or marker placement errors, rather than forming independent channel noise lacking physical connection.

[0165] This step processes the angle subspaces on the left and right sides separately to simulate the actual situation where the installation direction or anatomical calibration error of the rigid body on the left and right sides is not exactly the same.

[0166] S42, Global Scaling

[0167] Global scaling is used to simulate the differences in overall movement amplitude between different subjects or between different walking trials of the same subject.

[0168] In actual walking, different subjects may exhibit different overall amplitudes of motion due to differences in height, stride length, joint mobility, walking habits, or compensatory strategies. Even within the same functional state or disease category, the overall amplitude of their six-degree-of-freedom kinematic trajectories may vary. To simulate this variation in overall amplitude, a scaling factor is randomly sampled for each training sample to be enhanced.

[0169] ;

[0170] And apply the same scaling factor uniformly to all channels and all time points of this training sample:

[0171] ;

[0172] when When, the overall amplitude of the enhanced kinematic signal increases relative to the original signal; when At that time, the overall amplitude of the enhanced kinematic signal is reduced relative to the original signal.

[0173] Before and after global scaling, the peak occurrence time, period length and main waveform profile of each channel remain unchanged, and the original relative change relationship between each channel is also preserved. The main difference is that the overall amplitude of all kinematic signals changes in a consistent manner.

[0174] Since the same scaling factor applies to all channels simultaneously, this operation simulates the overall motion amplitude difference at the sample level, rather than the independent abnormal changes in a certain local time period or a certain channel.

[0175] S43, Amplitude Distortion

[0176] Amplitude distortion is used to simulate local amplitude fluctuations that occur in different time periods or different time phases during continuous walking of the same subject.

[0177] Global scaling uses a fixed scaling factor to change the overall amplitude of the complete sequence, while actual gait changes usually do not occur proportionally across all time points. For example, a subject may experience a slightly larger flexion amplitude during the swing phase of a step, while the movement changes less during the support phase; or the angle or translation amplitude may change gradually in a local time period due to stride adjustment, fatigue, or temporary compensation.

[0178] To simulate such local and continuous amplitude variations, for each channel Several amplitude control points are sampled, and interpolation is performed along the time direction to generate a smooth amplitude modulation curve with the same length as the original sequence. .

[0179] Subsequently, the smoothed amplitude modulation curve is applied to the corresponding channel at each time point:

[0180] ;

[0181] The smooth amplitude modulation curve changes continuously in the time direction, so the enhanced signal will not exhibit abrupt changes introduced by the enhancement operation itself between adjacent sampling points. Time intervals with larger curve values ​​correspond to local amplitude enhancement, while time intervals with smaller curve values ​​correspond to local amplitude reduction.

[0182] Before amplitude distortion, a certain channel maintains its original amplitude distribution in the complete sequence; after amplitude distortion, its main motion trend and periodic structure are still preserved, but the amplitude in different time intervals is amplified or reduced to different degrees according to the smooth curve.

[0183] Each channel can have a different smooth amplitude modulation curve, thus this operation can simulate different degrees of freedom responding differently to gait changes. At the same time, because the amplitude modulation curve is generated smoothly, the enhancement result is different from instantaneous jumps or discrete spikes, and is closer to the local amplitude changes generated during continuous motion.

[0184] S44, Time Distortion

[0185] Time warp is used to simulate the local time stretching or compression caused by changes in walking speed and the duration of different phases within the gait cycle.

[0186] Actual walking speed does not necessarily remain strictly constant throughout the data collection process. Subjects may increase or decrease their walking speed at certain times, and may also exhibit different time proportions between the support and swing phases. Such variations primarily alter the location and duration of kinematic events on the timeline, rather than significantly changing the amplitude of the movement trajectory itself.

[0187] To simulate the aforementioned time variation, several time offset control points were first sampled along the original time axis, and continuous time displacement curves were obtained through interpolation. Construct a non-uniform sampling index based on the continuous-time displacement curve:

[0188] ;

[0189] Subsequently, the original multi-channel time-series signal is resampled according to the non-uniform sampling index:

[0190] ;

[0191] when When corresponding to non-integer sampling positions, interpolation can be used to obtain the corresponding signal value; in specific implementation, the non-uniform sampling index is limited to the effective time range of the original multi-channel timing signal.

[0192] Regions where the time displacement curve changes slowly correspond to regions where the time progression of the original multi-channel timing signal is basically maintained; regions where the sampling index changes rapidly correspond to local time compression; and regions where the sampling index changes slowly correspond to local time stretching.

[0193] Before and after time warp, the number of channels and the length of the output sequence of the kinematic signal remain unchanged, but the positions of peaks, troughs and other kinematic events on the time axis are smoothly shifted, and the duration of the local gait phase also changes accordingly.

[0194] Preferably, the constructed non-uniform sampling index is used for all kinematic channels of the same sample. Instead of generating independent time indices for each channel, the same time mapping is used. This maintains the temporal correspondence between adduction / abduction, internal / external rotation, flexion / extension, and the three translational degrees of freedom, avoiding the disruption of the multi-degree-of-freedom kinematic structure at the same moment due to the independent misalignment of different channels.

[0195] Therefore, time warp alters the common time progression of multi-channel signals, but does not change the original synchronization relationship between different degrees of freedom.

[0196] S45, Random Bias

[0197] Random bias is used to simulate channel baseline offset caused by differences in initial anatomical calibration, coordinate system origin determination, or equipment zero-point setting.

[0198] In actual data acquisition, even if the dynamic changes of a joint's motion trajectory are completely consistent, there may be slight differences in the initial angles or initial translational positions obtained from different calibrations. For example, the same flexion-extension trajectory may be translated in the positive or negative angle direction as a whole, but its peak-to-valley difference, periodic structure, and temporal variation trend remain basically unchanged.

[0199] To simulate this type of baseline bias, for each channel Independently sample a random constant bias:

[0200] ;

[0201] This constant is then added to all time points of the corresponding channel: ;

[0202] Before random bias, the signal changes around the original calibration baseline; after random bias, the complete waveform shifts up or down along the numerical axis, but the distance between the peak and valley, the gait period, and the waveform profile remain unchanged.

[0203] Different channels can employ different random constant biases to simulate small deviations in anatomical calibration or zero-point calibration for different degrees of freedom. Random biasing reduces the model's dependence on a fixed absolute baseline, allowing it to learn more about the dynamic changes in kinematic signals.

[0204] S46, shaking

[0205] Jitter is used to simulate background noise, optical tracking errors, or random measurement fluctuations of the acquisition equipment.

[0206] Unlike random bias, which applies the same constant throughout the entire time series, jitter varies randomly at different time points. For the first... Each channel samples Gaussian random noise at each time point, and each channel is then superimposed with independently sampled Gaussian random noise.

[0207] ;

[0208] And superimpose it onto the original signal: ;

[0209] in, Indicates the first Noise intensity of each channel.

[0210] Before jittering, the signal is a kinematic trajectory after filtering and quality control; after jittering, the main periodic structure, motion amplitude trend and peak and valley positions of the signal are basically preserved, but small random fluctuations are added near local sampling points.

[0211] The jittering operation in this step is used to expose the model to similar signals at different noise levels during the training phase, reducing the model's dependence on individual sampling points or overly smoothed waveforms, thereby improving its tolerance to background noise from real equipment and slight measurement jitter.

[0212] The jitter adds random noise with a mean of zero, so its effect is mainly manifested as localized, subtle fluctuations, rather than causing the complete sequence to continuously shift in a certain direction. This differs from the random bias in step S35.

[0213] In this embodiment, the augmentation operations are randomly scheduled and combined. For each training sample, a predetermined execution probability is set for each of the six augmentation operations. Each time the constructed classification network model reads a training sample, it determines whether to implement the corresponding augmentation operation based on the corresponding execution probability.

[0214] Any augmentation operation can be applied alone or in combination with other augmentations. For example, crosstalk can be used to simulate angular coordinate system deflection, time warp can be used to simulate gait changes, and random bias or jitter can be added. This allows the same original training sample to form different reasonable kinematic variants in different training rounds.

[0215] When multiple operations are executed in combination, each operation acts on different data attributes: crosstalk disturbance changes the component distribution relationship between the three rotational degrees of freedom; global scaling changes the overall motion amplitude of all time-series signals; amplitude distortion changes the local amplitude of a single channel in different time intervals; time distortion changes the position and duration of kinematic events on the time axis; random bias changes the overall baseline of each channel; jitter changes the local random fluctuations near each time sampling point.

[0216] Therefore, the six enhancement operations do not repeatedly process the same signal attribute, but simulate the acquisition changes from the aspects of coordinate space, amplitude space, time space, calibration baseline and measurement noise respectively.

[0217] The augmentation operation does not change the original training sample class labels, nor does it change the number of input channels or the required time length for the model. The augmented signal is used together with the original training samples for model training, enabling the model to learn a wider range of reasonable kinematic performance within the same class.

[0218] The validation and test data are not augmented as described above; instead, they are directly input into the trained model according to a predetermined data preprocessing procedure. This ensures that the validation and test results reflect the model's performance on real, un-augmented data, avoiding interference from random augmentation operations in the performance evaluation.

[0219] Step S5: Construct a multi-path feature fusion classification network (PACENet) model, and train the constructed classification network model using an expanded training sample set; use the trained classification network model to classify time-series signals to obtain the classification prediction results of knee joint function.

[0220] This embodiment designs an attention-based compensatory feature extraction network (PACENet) to classify multi-degree-of-freedom kinematic temporal signals of bilateral knee joints. PACENet includes three parallel feature extraction branches: a temporal feature extractor (TFE), a frequency domain feature extractor (SFE), and a kinematic feature extractor (KFE). The outputs of the three feature extraction branches are fused along the channel dimension and then fed into the classification head.

[0221] S51, Temporal Feature Extractor (TFE)

[0222] The TFE design is inspired by observations in knee dysfunction assessment: knee dysfunction often manifests as bilateral asymmetry and compensatory gait patterns, while unilateral local abnormalities provide complementary information. Therefore, the TFE includes a unilateral feature extractor (UFE) and a compensatory feature extractor (CFE).

[0223] A one-sided feature extractor is used to extract one-sided local anomaly information. The two-sided input sequence is then used. Divided into independent left and right side sequences (each) Each side is processed separately using a Transformer architecture with shared weights. For each side, a 2D convolution is first applied to generate a patch token, then positional encoding is added, and then the token is fed into a multi-head self-attention (MHA) module. Finally, the MHA output is passed through a feedforward network (FFN) and an attention pooling layer to obtain the 1 / 2-sided feature. .

[0224] single-sided patch token Attention calculation is as follows:

[0225] ;

[0226] in For the number of tokens, For the feature dimension, , , .

[0227] In this embodiment, UFE uses shared weights to process the left and right limbs, ensuring that the features of the left and right knee joints are comparable in the same representation space, thereby supporting subsequent asymmetry analysis.

[0228] CFE is used to model bilateral compensatory relationships. Taking the position-encoded patch token output by UFE as input, CFE uses two cross-attention modules to model the dependency between the left and right knee joints. In each cross-attention module, the patch token of the target limb... Projection as a query, for the patch token of the lateral limb Projection into Key and Value:

[0229] ;

[0230] Through this mechanism, the representation of the target knee joint can adaptively integrate information from the contralateral knee joint, thereby capturing the asymmetry and compensatory features of both knee joints. The results are then processed using FFN and attention pooling to obtain the compensatory features. .

[0231] In the feature fusion stage, PACENet outputs CFE. The main branch will output the UFE. As a residual component, it is derived from learnable parameters. Controlling its contribution:

[0232] ;

[0233] This design allows the network to be trained in the early stages ( It mainly relies on CFE representations, and UFE representations are gradually and adaptively injected as training progresses. This is the final output feature of TFE.

[0234] In this embodiment, the combination of the dual-branch structure of UFE and CFE, and the learnable residual gating γ, mimics the knee joint dysfunction assessment logic of researchers simultaneously observing unilateral abnormalities and bilateral compensations. The shared-weight UFE ensures comparability between the left and right knee joints, the cross-attention CFE explicitly models the compensation relationship, and the γ gating achieves adaptive balance of the two information streams.

[0235] S52, Frequency Domain Feature Extractor (SFE)

[0236] Since TFE primarily models local temporal dependencies, its ability to explicitly characterize gait rhythm changes and spectral energy redistribution related to knee joint impairment is limited. SFE is used to supplement the frequency domain representation. Considering that the bilateral input sequences are finite-length discrete signals, directly applying the Fast Fourier Transform (FFT) will introduce spectral leakage due to boundary discontinuities, causing the dominant frequency energy to spread to adjacent frequency points, which may mask weak frequency components that still have discriminative power.

[0237] To suppress spectral leakage, SFE first processes the input sequence. The applied length is Hamming window:

[0238] ;

[0239] Apply a real-valued fast Fourier transform (rFFT) to the windowed signal and calculate the logarithmic amplitude spectrum:

[0240] ;

[0241] in To prevent Numerical instability of small constants. The resulting global logarithmic spectrum, after being flattened along the channel dimension, is fed into a projection network consisting of a Linear layer, LayerNorm, and GELU activation to obtain frequency domain features.

[0242] SFE employs a specific pipeline combining Hamming window, rFFT, and logarithmic spectrum to capture relatively stable gait rhythmic features across centers. Ablation experiments show that removing SFE causes the F1 score on the external test set to drop from 0.9156 to 0.7849, demonstrating that frequency domain modeling is crucial for cross-center generalization.

[0243] S53, Kinematic Feature Extractor (KFE)

[0244] While TFE and SFE provide expressive nonlinear gait dynamics representations, purely data-driven implicit features fail to accurately capture the functional descriptors commonly used in knee function assessment. KFE is used to supplement the learned representations by explicitly encoding knee function information. Since gait frequency and cycle duration vary from person to person, a fixed window cannot be used to segment cycles, and abnormal cycles may exist in the subject data, affecting descriptor computation. Therefore, KFE performs the same adaptive gait cycle segmentation as step S3, as well as abnormal cycle removal based on gait cycle integrity and repetition cycle consistency.

[0245] For the reserved first One effective original cycle (length) ),set up For flexion and extension angle, For inside and outside corners, For anterior-posterior translation. Extract the knee joint function descriptors shown in Table 1 below:

[0246] Table 1

[0247]

[0248] The above features are averaged over all effective cycles for each limb, and the results for both limbs are concatenated to form the original feature vector. After normalization by a predefined scaling factor, the data is fed into a feature projection module consisting of Linear, BatchNorm, GELU, and Dropout, mapping it to a latent space of the same dimension as TFE / SFE. Dropout is used to reduce over-reliance on any single explicit feature and improve training stability.

[0249] KFE explicitly encodes the above six knee joint function descriptors into feature vectors and projects them into a deep feature space, aligning model decisions with knee joint dysfunction assessment experience and compensating for the lack of interpretability of purely data-driven representations.

[0250] S54, Fusion and Classification Head

[0251] In the final decision-making stage, PACENet concatenates the features extracted by TFE, SFE, and KFE along the channel dimension to form a high-dimensional feature vector, which is then fed into a feature fusion module consisting of three layers of multilayer perceptron (MLP) for dimensionality reduction and information aggregation. Finally, the classification head outputs three predictions: normal knee joint function, injury-related knee joint dysfunction, and degenerative knee joint dysfunction.

[0252] PACENet's three-way converged overall architecture is as follows: Figure 2 As shown; the UFE / CFE dual-branch structure and residual gating fusion mechanism of TFE are as follows: Figure 3 As shown.

[0253] In this embodiment, the constructed multi-path feature fusion classification network model is trained using a data-augmented training sample set; cross-entropy loss is used as the loss function, AdamW is selected as the optimizer, and the initial learning rate is set to... The weight decay was set to 0.01, and 5-fold cross-validation was used within the development set, with each fold selecting the model with the highest F1 score on the validation set.

[0254] The trained model classifies the multi-degree-of-freedom kinematic time-series signals of both knee joints to obtain a three-class output, namely, the model outputs three types of predicted probabilities: normal knee joint function, injury-related knee joint dysfunction, and degenerative knee joint dysfunction.

[0255] Based on the same inventive concept, this embodiment also provides a knee joint dysfunction classification and recognition system based on knee joint gait kinematics, implemented using the above-mentioned knee joint dysfunction classification and recognition method, including:

[0256] The data acquisition module collects multi-degree-of-freedom kinematic time-series signals of both knee joints. The kinematics of each knee joint is described by a multi-dimensional characterization that includes at least rotational and translational degrees of freedom.

[0257] The preprocessing module performs standardized preprocessing on the collected bilateral knee joint multi-degree-of-freedom kinematic time-series signals, and screens the kinematic range of the time-series signals according to the rotational and translational degrees of freedom to identify physically unreliable kinematic trajectories and remove bilateral knee joint multi-degree-of-freedom kinematic time-series signal samples with abnormal rotational or translational degrees of freedom.

[0258] The quality control module performs quality control on the time-series signals after kinematic range screening based on gait cycle integrity and repetition cycle consistency, and identifies incomplete cycles or morphologically abnormal cycles.

[0259] The data augmentation module performs data augmentation on the preprocessed and quality-controlled time-series signals to obtain augmented data, which is used to expand the training samples.

[0260] The data augmentation process involves simulating different types of reasonable acquisition perturbations based on the acquisition method and physical structure of the multi-degree-of-freedom kinematic time-series signal of the knee joint, and expanding the kinematic performance forms that may correspond to the same training sample while keeping the original training sample category labels unchanged.

[0261] A multi-path feature fusion classification network model is trained using an expanded training sample set. The trained classification network model is then used to classify time-series signals to obtain classification prediction results for knee joint function.

[0262] The specific implementation methods of the above modules are described in steps S1-S5, and will not be repeated here.

[0263] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A method for classifying and identifying knee joint dysfunction based on knee joint gait kinematics, characterized in that, Includes the following steps: S1. Acquire multi-degree-of-freedom kinematic time-series signals of bilateral knee joints. The kinematics of each knee joint is described by a multi-dimensional characterization that includes at least rotational and translational degrees of freedom. S2. Standardize and preprocess the collected bilateral knee joint multi-degree-of-freedom kinematic time-series signals, and screen the kinematic range of the time-series signals according to the rotational and translational degrees of freedom to identify physically unreliable kinematic trajectories and remove bilateral knee joint multi-degree-of-freedom kinematic time-series signal samples with abnormal rotational or translational degrees of freedom. S3. Perform quality control on the time-series signals after kinematic range screening based on gait cycle integrity and repetition cycle consistency to identify incomplete cycles or morphologically abnormal cycles. S4. Perform data augmentation on the preprocessed and quality-controlled time-series signal to obtain augmented data, which is used to expand the training samples. The data augmentation process involves simulating different types of reasonable acquisition perturbations based on the acquisition method and physical structure of the multi-degree-of-freedom kinematic time-series signal of the knee joint, and expanding the kinematic performance forms that may correspond to the same training sample while keeping the original training sample category labels unchanged. S5. Construct a multi-path feature fusion classification network model, and train the constructed classification network model using an expanded training sample set; use the trained classification network model to classify time-series signals to obtain the classification prediction results of knee joint function.

2. The method for classifying and identifying knee joint dysfunction according to claim 1, characterized in that, The degrees of freedom of motion in the multi-degree-of-freedom kinematic timing signal of the knee joint include adduction and abduction, internal rotation and external rotation, flexion and extension, anterior-posterior translation, superior-inferior translation, and lateral-lateral translation; among them, adduction and abduction, internal rotation and external rotation, and flexion and extension are rotational degrees of freedom, while anterior-posterior translation, superior-inferior translation, and lateral-lateral translation are translational degrees of freedom.

3. The method for classifying and identifying knee joint dysfunction according to claim 1, characterized in that, The kinematic range screening in step S2 includes: Rotational degree of freedom anomaly screening involves statistically analyzing the absolute values ​​of the angles of the rotational degrees of freedom of the time-series signal throughout the entire acquisition time sequence and calculating the maximum absolute angle. The calculated maximum absolute angle is then compared with the preset upper limit threshold of the corresponding motion degree of freedom to determine whether the rotational degree of freedom of the knee joint on the corresponding side of the time-series signal is abnormal. Translational degree of freedom anomaly screening involves calculating the difference between the maximum and minimum values ​​of the translational degree of freedom of the time-series signal throughout the entire acquisition time sequence to obtain the translational motion range; comparing the calculated translational motion range with the preset translational range threshold of the corresponding motion degree of freedom to determine whether the knee joint translational degree of freedom on the corresponding side of the time-series signal is abnormal.

4. The method for classifying and identifying knee joint dysfunction according to claim 1, characterized in that, Step S3 first estimates the gait cycle based on the periodic characteristics of the flexion-extension angle signal in the multi-degree-of-freedom kinematic time-series signal of each knee joint, then determines the cycle boundary based on the estimation results of the gait cycle, and selects representative complete cycles through cycle time normalization and morphological similarity evaluation to achieve cycle segmentation. Step S3 involves performing gait cycle estimation and cycle segmentation on the multi-degree-of-freedom kinematic timing signals of the left and right knee joints, and then completing the structural pairing of the bilateral knee joint cycles based on the effective number of cycles on both knee joints.

5. The method for classifying and identifying knee joint dysfunction according to claim 4, characterized in that, Step S3 includes: S31. Utilize the repetitive variation pattern of the complete flexion-extension angle signal to estimate the gait period based on the autocorrelation function of the flexion-extension angle signal; S32. Based on the gait cycle estimate obtained in step S31, set the peak significance condition and the minimum distance between adjacent peaks, and perform peak detection on the flexion-extension angle signal to determine the cycle boundary. S33. Perform linear interpolation along the time dimension on the estimated gait cycle to unify it into a normalized cycle length; then perform linear interpolation on multiple degrees of freedom channels of the estimated gait cycle respectively, so that the first time point and the last time point correspond to the start and end points of the normalized cycle respectively, and the remaining time points are mapped according to the relative cycle position to obtain the time-normalized gait cycle. S34. Based on the gait cycles after time normalization of the left and right knee joints, count the number of effective normalized cycles extracted from the left and right knee joints, and take the smaller value of the two as the number of normalized paired cycles for bilateral knee joint modeling. S35. Evaluate the morphological similarity between normalized paired cycles of different knee joints of the same subject; based on morphological similarity, screen abnormal cycles to obtain the actual number of complete cycles of both knee joints retained by the subject; if the actual number of complete cycles of both knee joints retained is greater than the preset minimum number of effective cycles, the subject is deemed to have passed the quality control of gait cycle integrity and morphological consistency.

6. The method for classifying and identifying knee joint dysfunction according to claim 5, characterized in that, Step S31 also estimates the gait period based on the number of oscillation peaks in the flexion-extension angle signal; When at least two valid oscillation peaks are detected, the minimum value between the period estimation result obtained by the autocorrelation function and the period estimation result obtained based on the number of oscillation peaks is taken as the final adaptive gait period estimate. When fewer than two oscillation peaks are detected, the period estimation result obtained through the autocorrelation function will be used as the final adaptive gait period estimate.

7. The method for classifying and identifying knee joint dysfunction according to claim 6, characterized in that, Step S31 estimates the gait period based on the autocorrelation function of the flexion-extension angle signal, including: Calculate the average value of the flexion-extension angle signal at all time sampling points, and subtract the average value from the original flexion-extension angle signal to obtain the mean-reduced centered signal; Calculate the autocorrelation function for different candidate time delays on the centered flexion-extension angle signal; Within the candidate delay range, the delay corresponding to the maximum value of the autocorrelation function is taken as the period estimation result based on the overall periodicity. Step S31 estimates the gait period based on the number of oscillation peaks in the flexion-extension angle signal, including: Calculate the dynamic range of the flexion-extension angle signal; A peak significance threshold is set based on the dynamic range of the flexion-extension angle signal, and candidate oscillation peaks are detected; When at least two oscillation peaks are detected, a period estimation result based on peak count is calculated according to the length of the complete flexion-extension angle signal and the number of oscillation peaks.

8. The method for classifying and identifying knee joint dysfunction according to claim 1, characterized in that, The data augmentation process in step S4 includes at least one of crosstalk perturbation, global scaling, amplitude warping, time warping, random bias, and jitter. Crosstalk, global scaling, amplitude distortion, time distortion, random bias, and jitter respectively alter the channel coupling relationship, overall amplitude, local amplitude, time progression, baseline position, and local random fluctuations of a timing signal. For each training sample, a predetermined execution probability is set for different data augmentation processes; each time the constructed classification network model reads a training sample, it determines whether to implement the corresponding data augmentation process based on the corresponding execution probability.

9. The method for classifying and identifying knee joint dysfunction according to claim 1, characterized in that, The multi-path feature fusion classification network model includes three parallel feature extraction branches: a temporal feature extractor, a frequency domain feature extractor, and a kinematic feature extractor. The outputs of the three feature extraction branches are fused along the channel dimension and then fed into the classification head. The trained multi-path feature fusion classification network model outputs three predicted probabilities: normal knee joint function, injury-related knee joint dysfunction, and degenerative knee joint dysfunction.

10. A knee joint dysfunction classification and identification system based on knee joint gait kinematics, characterized in that, The method for classifying and identifying knee joint dysfunction as described in claim 1 is used, comprising: The data acquisition module collects multi-degree-of-freedom kinematic time-series signals of both knee joints. The kinematics of each knee joint is described by a multi-dimensional characterization that includes at least rotational and translational degrees of freedom. The preprocessing module performs standardized preprocessing on the collected bilateral knee joint multi-degree-of-freedom kinematic time-series signals, and screens the kinematic range of the time-series signals according to the rotational and translational degrees of freedom to identify physically unreliable kinematic trajectories and remove bilateral knee joint multi-degree-of-freedom kinematic time-series signal samples with abnormal rotational or translational degrees of freedom. The quality control module performs quality control on the time-series signals after kinematic range screening based on gait cycle integrity and repetition cycle consistency, and identifies incomplete cycles or morphologically abnormal cycles. The data augmentation module performs data augmentation on the preprocessed and quality-controlled time-series signals to obtain augmented data, which is used to expand the training samples. The data augmentation process involves simulating different types of reasonable acquisition perturbations based on the acquisition method and physical structure of the multi-degree-of-freedom kinematic time-series signal of the knee joint, and expanding the kinematic performance forms that may correspond to the same training sample while keeping the original training sample category labels unchanged. A multi-path feature fusion classification network model is trained using an expanded training sample set. The trained classification network model is then used to classify time-series signals to obtain classification prediction results for knee joint function.

Images