Musculoskeletal disorder risk assessment and early warning method and system based on multi-source data

Abstract

The application provides a kind of muscle and skeleton disease risk assessment and early warning method and system based on multi-source data, it is related to muscle, skeleton assessment and early warning technical field, the application is by collecting joint angle, electromyography and foot bottom load data, constructs time sequence consistent joint feature vector sequence, effectively make up the limitation of traditional single monitoring means, by K means clustering to joint angle sequence, form standard posture set with physiological significance, and construct transition edge weight based on angle change, muscle activation and load fluctuation weighted calculation, realize the physiological cost quantization expression of different posture transfer path, it is convenient to identify risk transition area, by calculating the deviation between latent variable prediction value and true value, sudden transition behavior in the movement process can be captured, and combined with its edge weight intensity in directed graph, by considering transition intensity and behavior abnormality degree simultaneously, the muscle and skeleton disease risk is effectively warned.

Description

Technical Field

[0001] This invention relates to the field of muscle and bone assessment and early warning technology, specifically to a method and system for musculoskeletal disease risk assessment and early warning based on multi-source data. Background Technology

[0002] In recent years, with the development of intelligent rehabilitation, sports health management, and industrial ergonomics, higher demands have been placed on the dynamic monitoring and risk warning of key joints in the lower limbs. These joints bear high-frequency physiological load changes during human movement and are high-incidence sites for many musculoskeletal disorders. Especially in high-intensity labor, rehabilitation training, or sports competition scenarios, individuals are highly susceptible to irreversible damage due to sudden movements, poor posture, or fatigue accumulation. Therefore, early detection and early warning of the complex physiological loads borne by the lower limb joints during exercise, labor, and rehabilitation, and real-time assessment of their movement status are of great significance for preventing musculoskeletal injuries.

[0003] Commonly used methods in existing technologies include using inertial measurement units (IMUs) to monitor joint angles, utilizing surface electromyography (SEMG) to reflect muscle activation status, or estimating gait load using plantar pressure sensors. These methods can provide certain physiological parameter information in their respective dimensions and are widely used in clinical rehabilitation and sports science research. However, they are often used in isolation, lacking data fusion and correlation modeling, making it difficult to form a holistic profile of motor behavior.

[0004] Existing technologies still have significant shortcomings in several aspects: First, single-dimensional signals cannot fully characterize the complex evolution process of joint states, especially in complex dynamic movements, and cannot effectively reflect the coupling relationship between different physiological variables; Second, existing methods generally lack the ability to model the process of motion state "transition" (such as abrupt changes in posture or abnormal movements), resulting in a lag or even omission in the identification of potential anomalies.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for risk assessment and early warning of musculoskeletal disorders based on multi-source data, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for musculoskeletal disease risk assessment and early warning based on multi-source data, comprising the following steps:

[0009] The multi-channel data of the target joint are continuously acquired to form a joint feature vector sequence. The multi-channel data includes joint angle data corresponding to the timestamp, electromyographic signal data of the muscle surface associated with the target joint, and corresponding joint load data. The joint feature vector sequence includes joint angle sequence, electromyographic signal sequence, and joint load sequence.

[0010] K-means clustering was performed on the joint angle sequence to divide all joint angle data into multiple standard poses. Based on the joint angle data, electromyographic signal data and corresponding joint load data within each standard pose, edge weights between any two standard poses were constructed, and a directed graph network representing the behavioral evolution path was constructed according to the collection order.

[0011] The joint feature vector sequence is input into the variational autoencoder model for feature compression to obtain a low-dimensional latent variable representation sequence. The latent variable representation sequence is then input into the time series prediction model, and a gated recurrent unit is used for time series modeling. The latent variable representation of the next sampling point is predicted based on the latent variable representation of the previous sampling point in the latent variable representation sequence.

[0012] Calculate the deviation between each actual latent variable and the predicted latent variable. Based on the deviation, determine whether there is abnormal transition behavior between the multi-channel data corresponding to adjacent sampling points. Map the abnormal transition behavior to a directed graph network, count the cumulative frequency and edge weight distribution of the abnormal transition behavior, classify the risk level, and output the corresponding musculoskeletal disease early warning signal.

[0013] Furthermore, the target joint is one of the hip joint, knee joint, or ankle joint in the lower limb joints, and the joint load data is the average pressure on the soles of both feet;

[0014] When collecting multi-channel data of the target joint, multiple sampling points are set up. Each sampling point synchronously collects joint angle data, electromyographic signal data, and joint load data, and the time interval between adjacent sampling points is the same.

[0015] Furthermore, the logic for K-means clustering of the joint angle sequence is as follows:

[0016] Determine the maximum range of motion of the target joint, divide the maximum range of motion into K angle ranges corresponding to standard poses, and use the median of the angles within each angle range as the initial cluster center;

[0017] Calculate the Euclidean distance between each joint angle data and each initialized cluster center, assign the corresponding joint angle data to the nearest cluster center, and update each cluster center to the mean of all angle values ​​in its cluster.

[0018] Repeatedly calculate the Euclidean distance between each joint angle data and each updated cluster center, and reassign the corresponding joint angle data to the nearest cluster center until the cluster centers converge, that is, the cluster centers no longer change.

[0019] Map the standard pose corresponding to the cluster of each joint angle data to the standard pose corresponding to each joint angle data.

[0020] Furthermore, the method for constructing the edge weights between any two standard poses is as follows:

[0021] Calculate the average joint angles, root mean square values ​​of all electromyography (EMG) signals, and average joint loads within each cluster corresponding to the two standard postures. Also calculate the differences in joint angles, EMG signals, and joint loads between the two clusters. The formulas used are as follows:

[0022]

[0023] in, , and These represent the standard postures as follows: and The differences in joint angles, electromyographic signals, and joint loads between corresponding clusters at that time. and These represent the standard postures as follows: and The average joint angle corresponding to the cluster. and These represent the standard postures as follows: and The average joint load corresponding to the cluster at that time. and These represent the standard postures as follows: and The first corresponding cluster in the time The root mean square value of the electromyographic signals on the associated muscle surface. and All of these represent the standard posture number. and These represent the muscle number associated with the target joint and the total number of associated muscles, respectively.

[0024] The weighted average of the differences is used to construct the edge weights between two standard poses, based on the following formula:

[0025]

[0026] in, Indicates standard posture To standard posture Border rights during a jump , and These represent angle weights, signal weights, and load weights, respectively. ,and .

[0027] Furthermore, the method for constructing a directed graph network representing the behavioral evolution path is as follows:

[0028] The joint angle data, electromyographic signal data, and corresponding joint load data at each sampling point are taken as a time node. All time nodes are sorted according to the sampling timestamps corresponding to the sampling points. The direction of the directed graph network is the transition from the previous time node to the next time node. The edge weight of two adjacent time nodes is the edge weight when transitioning from the standard posture corresponding to the previous time node to the standard posture corresponding to the next time node.

[0029] Furthermore, the joint feature vector sequence is:

[0030]

[0031] in, For the first The joint feature vector of each sampling point and They represent the first Joint angle data and joint load data at each sampling point Indicates the first In the sampling points, the first Electromyographic signal data of a related muscle surface, This indicates the total number of muscles associated with the target joint. Indicates the total number of sampling points;

[0032] The method for obtaining a low-dimensional latent variable representation of a sequence is as follows:

[0033] The joint feature vector of each sampling point is input into the trained variational autoencoder, which outputs the mean vector and log-variance vector of the latent variables. The latent variable representation for each sampling point is obtained by sampling from the latent variable distribution using a reparameterization technique, based on the following formula:

[0034]

[0035] in, and They represent the first The latent variable representation and mean vector of each sampling point , Represents the log-variance vector. Standard normal noise, i.e. It follows a multidimensional Gaussian distribution with a mean of 0 and a covariance matrix of identity.

[0036] A low-dimensional latent variable representation sequence is formed according to the timestamps corresponding to the sampling points. The latent variable representation sequence is as follows: The method for predicting the latent variable representation of the next sampling point based on the latent variable representation of the previous sampling point in the latent variable representation sequence is as follows:

[0037] The trained gated recurrent unit is used to process the latent variable sequence. For time-series modeling, the latent variable representation of each sampling point is used as input. The initial hidden state of the gated loop unit is 0. It receives the latent variable representation of the current sampling point and the hidden state of the previous sampling point, and calculates the hidden state of the current sampling point through the gating mechanism.

[0038] Based on the hidden state of the current sampling point, a linear layer is used to predict the latent variable representation of the next sampling point. The underlying logic is as follows:

[0039]

[0040] in, Indicates the predicted first The latent variable representation predicted from each sampling point, Indicates the first The hidden state of each sampling point and These represent the learning weights and biases of the gated recurrent unit, respectively.

[0041] Furthermore, the deviation between the predicted latent variable representation and the true latent variable representation for each sampling point is calculated using the following formula:

[0042]

[0043] in, Indicates the first The deviation between the predicted latent variable representation and the true latent variable representation at each sampling point and They represent the first The predicted latent variable representation and the true latent variable for each sampling point, when When, determine the first The sampling point and the first There are abnormal transition behaviors between the multi-channel data of each sampling point. This indicates the deviation threshold.

[0044] Furthermore, the method for outputting corresponding musculoskeletal disease warning signals is as follows:

[0045] Abnormal transition behaviors are mapped onto a directed graph network. The edge weights corresponding to each abnormal transition behavior are obtained, the frequency of abnormal transition behaviors is counted, and the transition risk state of the target joint during the observation process is calculated by combining the deviation corresponding to each abnormal transition behavior. The formula used is as follows:

[0046]

[0047] in, Indicates a transition risk status. This indicates the total number of sampling points. That is, the number of all edge weights. and They represent the first The edge weights and biases corresponding to each anomalous transition behavior. This represents the total number of all abnormal jump behaviors;

[0048] like This indicates a low-risk warning, suggesting occasional abnormal transitions and low edge weights. A medium-risk warning was issued, indicating frequent abnormal transitions. The system issues a high-risk warning indicating multiple abnormal jumps, occurring on high-weight paths. It issued an extremely high-risk warning that required immediate intervention.

[0049] The present invention also provides a musculoskeletal disease risk assessment and early warning system based on multi-source data. The system is used to execute the above-described method for musculoskeletal disease risk assessment and early warning based on multi-source data, comprising:

[0050] The feature acquisition module is used to continuously acquire multi-channel data of the target joint to form a joint feature vector sequence. The multi-channel data includes joint angle data corresponding to the timestamp, electromyographic signal data of the muscle surface associated with the target joint, and corresponding joint load data. The joint feature vector sequence includes joint angle sequence, electromyographic signal sequence, and joint load sequence.

[0051] The network construction module is used to perform K-means clustering on the joint angle sequence, divide all joint angle data into multiple standard poses, construct edge weights between any two standard poses based on the joint angle data, electromyographic signal data and corresponding joint load data within each standard pose, and construct a directed graph network representing the behavioral evolution path according to the collection order.

[0052] The variable prediction module is used to input the joint feature vector sequence into the variational autoencoder model for feature compression to obtain a low-dimensional latent variable representation sequence. The latent variable representation sequence is then input into the time series prediction model, and a gated recurrent unit is used for time series modeling. Based on the latent variable representation of the previous sampling point in the latent variable representation sequence, the latent variable representation of the next sampling point is predicted.

[0053] The risk assessment module is used to calculate the deviation between each actual latent variable and the predicted latent variable, determine whether there is abnormal transition behavior between the multi-channel data corresponding to adjacent sampling points based on the deviation, map the abnormal transition behavior to a directed graph network, count the cumulative frequency and edge weight distribution of abnormal transition behavior and classify the risk level, and output the corresponding musculoskeletal disease early warning signal.

[0054] Compared with the prior art, the beneficial effects of the present invention are:

[0055] This invention effectively overcomes the limitations of traditional single monitoring methods by collecting joint angle, electromyography, and plantar load data to construct a time-consistent joint feature vector sequence, providing a data foundation for subsequent multimodal joint modeling. Secondly, by performing K-means clustering on the joint angle sequence, a set of physiologically meaningful standard postures is formed, and transition edge weights based on angle changes, muscle activation, and load fluctuations are constructed. This enables a quantitative expression of the physiological costs of different posture transition paths, facilitating the identification of risky transition regions.

[0056] Variational autoencoders are used to compress high-dimensional joint features to obtain low-dimensional representations of behavioral states in the latent variable space. A gated recurrent unit is then used to predict and model the latent variable sequence, establishing a baseline trajectory for the temporal evolution of individual behavior. Based on this, the deviation between the predicted and actual latent variable values ​​is calculated, which can capture sudden transition behaviors during movement. Combined with the edge weight strength in the directed graph, effective early warning of musculoskeletal disease risk is achieved by simultaneously considering the transition strength (edge ​​weight) and the degree of behavioral abnormality (latent variable deviation). Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the overall method flow of the present invention;

[0058] Figure 2 This is a schematic diagram of the joint feature vector sequence in this invention;

[0059] Figure 3 This is a schematic diagram of a directed graph network in this invention;

[0060] Figure 4 This is a schematic diagram of the overall method flow of the present invention;

[0061] Figure 5 This is a schematic diagram illustrating the generation of transition risk states in this invention. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0063] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0064] Example:

[0065] Please see Figures 1-4 The present invention provides a technical solution:

[0066] A method for musculoskeletal disease risk assessment and early warning based on multi-source data, comprising the following steps:

[0067] Step 1: Continuously acquire multi-channel data of the target joint to form a joint feature vector sequence. The multi-channel data includes joint angle data corresponding to the timestamp, electromyographic signal data of the muscle surface associated with the target joint, and corresponding joint load data. The joint feature vector sequence includes joint angle sequence, electromyographic signal sequence, and joint load sequence.

[0068] In this embodiment, the target joint is one of the hip, knee, or ankle joints in the lower limbs. The joint load data is the average pressure on the soles of both feet. The hip, knee, or ankle joints are selected as the monitoring objects because they are the core weight-bearing and movement hubs of the lower limbs, undertaking key mechanical and control functions in various high-frequency activities. These three types of joints are high-incidence sites for musculoskeletal disorders (such as arthritis, meniscus injury, chondromalacia patellae, etc.) in clinical practice, and are often caused by repetitive high loads, poor movement patterns, or muscle control disorders, resulting in strain and lesions. Therefore, continuous monitoring of the angle changes, muscle activation state, and mechanical load of key lower limb joints during movement is of great significance for early identification of disease risks and providing early warnings.

[0069] Furthermore, joint angle data refers to the relative angles between bone segments formed during the movement of a target joint. The joint angle in the knee joint is the flexion-extension angle between the thigh (femur) and the lower leg (tibia). The joint angle in the hip joint is the angle between the thigh and the trunk, covering flexion-extension, adduction-abduction, and rotation. The joint angle in the ankle joint is the angle between the lower leg and the foot, covering movements such as plantar flexion (toe down) and dorsiflexion (toe up). Joint angle data is obtained by wearing an inertial measurement unit on the thigh, lower leg, etc., collecting three-axis acceleration and angular velocity data, and combining it with existing attitude calculation algorithms (such as Madgwick and Kalman filtering) to calculate the bone segment attitude, thereby obtaining the joint angle.

[0070] The movement of the target joint is actively driven by the surrounding muscle groups. Therefore, monitoring the surface electromyography (EMG) signals of the muscles associated with the target joint can indirectly reflect the neural control load, muscle coordination, and fatigue level of that joint. The muscles associated with the knee joint include the quadriceps femoris, biceps femoris, and gastrocnemius. The muscles associated with the hip joint include the gluteus maximus, iliopsoas, and gluteus medius. The muscles associated with the ankle joint include the tibialis anterior, soleus, and gastrocnemius. The surface EMG signals are acquired by collecting them through surface electrode patches and connecting them to an EMG acquisition device.

[0071] In this embodiment, joint load data is used to indirectly estimate the vertical biomechanical load on the target joint by using the average pressure on the soles of both feet. The load on the lower limb joints mainly comes from the downward transmission of body weight, the upward transmission of ground reaction force, and the additional impact force caused by acceleration and inertia. These forces are transmitted upward through the soles of the feet to joints such as the ankle, knee, and hip, creating stress on the articular cartilage, ligaments, and tendons. The average pressure on the soles of both feet is obtained by collecting data through insole-type pressure sensors. The insole-type pressure sensors have multiple built-in pressure points and can calculate the average pressure.

[0072] When collecting multi-channel data of the target joint, multiple sampling points are set. Each sampling point synchronously collects joint angle data, electromyographic signal data, and joint load data, and the time interval between adjacent sampling points is the same. In order to achieve high timeliness and synchronization of the physiological load state of the target joint, multi-channel synchronous sampling is performed at a fixed time interval. Specifically, multiple sampling points are set during the acquisition process. The sampling points are discrete time nodes divided on the time axis by the system according to the set sampling frequency. Each sampling point corresponds to a specific acquisition time. Multiple sampling points are arranged in the order of timestamps to form a complete time series.

[0073] At each sampling point, multi-channel data of the target joint is simultaneously acquired, including: joint angle data at that moment, electromyographic signal data of multiple muscle surfaces associated with the target joint, and the average pressure value of the soles of both feet as joint load data. The sampling process has strict time alignment characteristics, that is, the above three types of data are acquired synchronously at the same sampling point to ensure that they have consistent time labels and can accurately reflect the motion posture, muscle activation state, and load of the target joint at the same moment. The time interval between sampling points is kept consistent, forming an equally spaced time series, which allows the system to perform dynamic modeling and behavioral evolution analysis based on a unified time scale. The multi-channel joint feature vector sequence constructed in this way has high temporal consistency and physiological parameter synergy, providing a data foundation and modeling guarantee for subsequent latent variable compression, sequence prediction, and anomaly detection.

[0074] Step 2: Perform K-means clustering on the joint angle sequence to divide all joint angle data into multiple standard poses. Construct edge weights between any two standard poses based on the joint angle data, electromyographic signal data and corresponding joint load data within each standard pose. Construct a directed graph network representing the behavioral evolution path according to the collection order.

[0075] In this embodiment, the logic for K-means clustering of the joint angle sequence is as follows:

[0076] Determine the maximum range of motion of the target joint, and divide the maximum range of motion into K angle ranges corresponding to standard postures. The maximum range of motion is the difference between the minimum angle value and the maximum angle value that the joint may reach during normal movement. Divide the angle range into equal intervals according to the set number of clusters K to form K initial angle sub-intervals corresponding to standard postures. The median of each angle sub-interval is selected as the initial cluster center of the standard posture, forming the initial set of K cluster centers.

[0077] For each sampling point's joint angle data, calculate its Euclidean distance to all K initial cluster centers, and assign the angle data to the cluster corresponding to the nearest cluster center. After one round of assignment, average the angle values ​​within each cluster to obtain a new cluster center. The new cluster center is the arithmetic mean of all angle values ​​in the current cluster, replacing the original center position.

[0078] The Euclidean distances between all angle data and the new cluster centers are recalculated, and the data is redistributed. Then, the mean values ​​of the angle values ​​within the new clusters are updated. The above data distribution and cluster center update process is executed iteratively until the variation values ​​between all cluster centers are lower than the preset convergence threshold or the positions of the cluster centers no longer change. At this point, the clustering process is considered complete, and the final K stable cluster centers are obtained.

[0079] After clustering, the system maps each joint angle data to its corresponding standard pose number according to the category label of its cluster. Thus, the original continuous joint angle sequence is transformed into a discrete standard pose label sequence, providing a structural foundation for subsequent directed graph modeling and behavioral evolution analysis based on pose transitions.

[0080] Specifically, to achieve standardized posture classification of the target joint's motion state, it is necessary to first determine the maximum range of motion of the target joint, and then perform K-means clustering on the collected joint angle data based on this angle range to obtain multiple standard posture categories. The classification method aims to ensure the stability of the initial clustering while making the posture interval division physiologically reasonable and numerically balanced, thereby improving the clustering convergence efficiency and the representativeness of the posture discretization results. Different lower limb joints in the human body have clear limits on their range of motion in terms of physiological structure. This range usually reflects the maximum flexion and extension trajectory that the joint can complete under active control. The system uses this maximum range of motion as the basis for classification, which helps to form a reasonable coverage of the complete range of motion in the initial stage of clustering, avoiding the initial cluster centers from clustering in a narrow angle interval, which would affect the final clustering effect.

[0081] Taking the knee joint as an example, the common range of motion is 0° (fully extended) to 135° (fully flexed). This 135° range is divided into K equal-width angle intervals. With the cluster number K=6, each interval is approximately 22.5° wide, forming six initial posture intervals: [0°, 22.5°], [22.5°, 45°], [45°, 67.5°], [67.5°, 90°], [90°, 112.5°], and [112.5°, 135°]. The corresponding initial cluster centers are set to the median of each interval, i.e., 11.25°, 33.75°, 56.25°, 78.75°, 101.25°, and 123.75°, respectively. [0°, 22.5°, 45°, 45°, 67.5°], [67.5°, 90°], [90°, 112.5°], and [112.5°, 135°]. The range [22.5°, 22.5°] corresponds to a fully extended knee or near-upright position, commonly seen in static standing or initial ready postures. The range [22.5°, 45°] corresponds to a slightly flexed knee posture, such as the support phase during starting or slow walking. The range [45°, 67.5°] corresponds to a moderately flexed knee posture, commonly seen when going up or down stairs or during the transition phase of a movement. The range [67.5°, 90°] corresponds to a standard squat posture or working squat. The range [90°, 112.5°] corresponds to a deep knee flexion phase, such as static load movements like sitting or half-kneeling. The range [112.5°, 135°] corresponds to the extreme knee flexion state, often occurring in ground-assisted movements or the moment of returning to standing up from a squat.

[0082] If the target joint is the hip joint, its common maximum range of motion is -30° to 120°, corresponding to a total flexion and extension angle of 150°. Under the clustering condition of K=5, each angle interval has a width of 30° and is divided into [-30°, 0°], [0°, 30°], [30°, 60°], [60°, 90°], and [90°, 120°], with cluster centers of -15°, 15°, 45°, 75°, and 105°, respectively, covering the entire process from maximum extension to maximum flexion. The [-30°, 0°] interval indicates that the hip joint is in a hyperextended state, commonly seen in brisk walking or leg swinging backward. The [0°, 30°] interval indicates that the hip joint is in an upright or slightly flexed state, typically seen in standing and gait support. The [30°, 60°] interval indicates moderate hip flexion, such as mid-walking or taking large strides. The [60°, 90°] interval indicates deep hip flexion, such as sitting down or lifting the leg while climbing stairs. The [90°, 120°] interval indicates extreme hip flexion, such as knee-hugging or deep sitting postures.

[0083] For the ankle joint, the range of motion for dorsiflexion and plantarflexion is generally -20° to 45°, totaling about 65°. This can be divided into equal-width intervals under K=4 conditions, with each posture covering approximately 16.25°. The corresponding cluster centers are the median values ​​of each interval, such as -10°, 8.125°, 24.375°, and 38.75°. The interval [-20°, 0°] indicates that the ankle joint is in a dorsiflexion state, commonly seen when the foot swings off the ground or when landing and cushioning. The interval [0°, 16.25°] indicates a neutral or mild plantarflexion state, seen when standing still or walking slowly. The interval [16.25°, 32.5°] indicates moderate plantarflexion, such as during the stepping or extension phase before jumping. The interval [32.5°, 45°] indicates strong plantarflexion, often occurring during vigorous movements such as jumping off the ground and toe extension.

[0084] Furthermore, the method for constructing the edge weights between any two standard poses is as follows:

[0085] Calculate the average joint angles, root mean square values ​​of all electromyography (EMG) signals, and average joint loads within each cluster corresponding to the two standard postures. Also calculate the differences in joint angles, EMG signals, and joint loads between the two clusters. The formulas used are as follows:

[0086]

[0087] in, , and These represent the standard postures as follows: and The differences in joint angles, electromyographic signals, and joint loads between corresponding clusters at that time. and These represent the standard postures as follows: and The average joint angle corresponding to the cluster. and These represent the standard postures as follows: and The average joint load corresponding to the cluster at that time. and These represent the standard postures as follows: and The first corresponding cluster in the time The root mean square value of the electromyographic signals on the associated muscle surface. and All of these represent the standard posture number. and These represent the muscle number associated with the target joint and the total number of associated muscles, respectively.

[0088] Furthermore, in order to achieve a quantitative description of the transition behavior between standard postures, it is necessary to construct edge weights between any two standard postures. These edge weights are used to measure the intensity of the state transition between the two standard postures, comprehensively reflecting the degree of angle change of the target joint, the difference in activation of the muscles involved, and the fluctuation of the joint load during the transition. As an attribute of the connection path between standard postures, the edge weights are an important foundation for the subsequent construction of directed graph network structures, identification of abnormal transition behaviors, and individualized risk scoring.

[0089] The process of constructing edge weights relies on statistical analysis of the data in each standard pose cluster. Specifically, it first involves considering two standard poses to be compared. and Data samples from all sampling points within each cluster were extracted, and the average values ​​of three key physiological indicators were calculated based on these samples: the average joint angle, the root mean square value of electromyography (EMG) of the muscle groups associated with the target joint, and the average plantar pressure as the average joint load. The average joint angle represents the range of motion corresponding to the posture, the average joint load reflects the mechanical load of the posture in steady state, and the root mean square value of EMG is used to represent the average activation level of the muscles mainly involved in control and stability in that posture. Assuming the standard posture is... There are G sampling points in the cluster, and the k-th associated muscle is in the . The signal at each sampling point is Then, in the standard posture, this muscle signal... The root mean square value is The calculation formula used is as follows: 。

[0090] In this embodiment, by calculating the differences in joint angles, electromyographic signals, and joint loads, the physiological state changes between standard postures can be comprehensively characterized, and further reflected on the changes in the range of motion, physiological energy consumption, and mechanical load when an individual performs a movement transition. The magnitude of the above differences directly reflects the complexity of the movement transition and the potential risk impact on the musculoskeletal system.

[0091] Joint angle difference values ​​reflect the degree of change in joint range of motion during an individual's transition from one standard posture to another. A larger difference value indicates a more drastic change in joint range of motion during the transition, requiring the joint to rapidly complete complex movements such as large-amplitude flexion, extension, or rotation. This may lead to increased soft tissue stretching and joint stress concentration, thereby increasing the risk of fatigue and injury to the musculoskeletal system. Electromyographic (EMG) signal difference values ​​reflect the activation changes of muscles associated with the target joint during the transition. The root mean square (RMS) value of the EMG signal is an effective indicator of muscle activity intensity; a larger difference value indicates that the muscles need to significantly increase or decrease their contraction intensity to coordinate with the joint in completing the posture change during the transition. Such significant changes in muscle activation are often accompanied by higher energy consumption and muscle fatigue accumulation, and may also reflect muscle control imbalance or compensation during the transition. These values ​​are important indicators for evaluating changes in muscle load and fatigue risk.

[0092] Joint load variation directly reflects the changes in external mechanical load on the joints during a jump. Joint load is usually reflected by the average pressure on the soles of both feet. The larger the variation, the more drastic the changes in ground reaction force or body weight load during the movement, which may lead to a sharp increase in instantaneous force on the joints. In this case, the mechanical impact on the articular cartilage, ligaments, and tendons increases, further exacerbating the risk of wear and damage to joint tissues. The larger the variation in joint angle, electromyographic signal, and joint load, the more drastic the changes in exercise intensity, muscle load, and mechanical stress experienced by the individual when jumping from the current standard posture to the target standard posture. Jumps with large variation values ​​often have a higher physiological burden and potential injury risk, and are high-risk moments for musculoskeletal disorders. Therefore, the boundary weights calculated based on these variation values ​​can effectively characterize the overall risk level of the jump process.

[0093] The weighted average of the differences is used to construct the edge weights between two standard poses, based on the following formula:

[0094]

[0095] in, Indicates standard posture To standard posture Border rights during a jump , and These represent angle weights, signal weights, and load weights, respectively. ,and .

[0096] Furthermore, constructing edge weights between standard postures aims to quantify the transition strength from one standard posture to another, and the impact of joint movements, muscle activation, and load changes involved in this transition on the individual's physiological system. By weighted averaging of joint angles, electromyographic signals, and joint load differences, the transition strength and physiological load level between standard postures can be scientifically measured. This allows for the assignment of a numerical "edge weight" to each edge in the constructed directed graph network, used for subsequent abnormal transition analysis, risk assessment, and early warning mechanisms. , and These represent the importance weights of joint angle differences, electromyographic signal differences, and joint load differences in the edge weight calculation. Changes in joint angle directly reflect changes in range of motion; larger angle changes usually mean that the joint needs to bear a higher physiological load. Therefore, angle differences have a higher weight in the edge weight calculation. This indicates that angle changes have a higher priority when assessing transition strength. Changes in electromyography (EMG) signals reflect changes in muscle activation and control ability. Larger EMG signal differences usually mean heavier muscle load and can also reflect possible muscle fatigue or imbalance during movement. Therefore, EMG signals have a certain weight in the weighting. Differences in joint load reflect changes in the mechanical load borne by the joint. Although load differences are very important in some movements, compared with changes in angle and EMG, changes in joint load are usually more affected by movement mode and external environment. Therefore, the weight of load differences is relatively small.

[0097] Based on practical applications and data analysis, weights , and The settings can be adjusted according to the characteristics of the target joint and the needs of the research object. In this embodiment, it can be set to... , , This setup ensures that joint angle changes dominate, while electromyographic signals and load changes moderately influence the final edge weight calculation, making it suitable for most physiological activity monitoring scenarios.

[0098] The magnitude of the edge weight directly reflects the physiological load of the transition between standard postures. The larger the edge weight, the greater the change in joint angle, the larger the range of motion, the larger the change in electromyographic signal, and the higher the degree of muscle activation during the transition. This may indicate a greater risk of fatigue accumulation or imbalance. The larger change in joint load indicates that the mechanical load on the joint changes drastically during the transition, which may increase the risk of joint injury.

[0099] Furthermore, the method for constructing a directed graph network representing the behavioral evolution path is as follows:

[0100] The joint angle data, electromyographic signal data, and corresponding joint load data at each sampling point are taken as a time node. All time nodes are sorted according to the sampling timestamps corresponding to the sampling points. The direction of the directed graph network is the transition from the previous time node to the next time node. The edge weight of two adjacent time nodes is the edge weight when transitioning from the standard posture corresponding to the previous time node to the standard posture corresponding to the next time node.

[0101] To accurately represent the behavioral evolution path of the target joint during movement and to construct a directed graph network based on the transitions of the standard posture, the system first combines and processes the joint angle data, electromyographic signal data, and joint load data at each sampling point. Each sampling point represents a time node, signifying the state of the target joint at a specific moment. These time nodes are ordered according to the timestamps of the sampling points, forming a strictly ordered time series. In this time series, the data from each sampling point constitutes a node in the graph, containing the joint angle, electromyographic signal intensity, and joint load data of the target joint at that moment. To form the directed graph network, directed edges are established in the order from one time node to the next, with each edge representing a transition from the standard posture. To standard posture The transition process is shown in the figure. The edge weights in the figure reflect the "physiological transition intensity" between one time node and the next time node, that is, the edge weight value between the standard posture corresponding to the previous time node and the standard posture corresponding to the next time node.

[0102] In this directed graph network, time nodes represent the individual's posture state at each time point, while edges represent the transition from one posture to another. The edge weight between two adjacent time nodes is calculated by weighting the differences in angle, electromyography, and load between the previous and subsequent standard postures. This edge weight reflects the amplitude of the target joint's posture change, the change in muscle activation, and the degree of load change on the joint within a specific time period. A larger edge weight indicates that the target joint experienced a larger amplitude of movement or load change during the transition, resulting in a heavier physiological burden and potentially higher risks. In this way, the movement process of the target joint is transformed into a directed graph network, and the edge weights between nodes quantify the individual's behavioral evolution trajectory through the posture transitions during movement. This directed graph network provides the foundation for subsequent abnormal transition detection, risk assessment, and early warning signal output, enabling timely identification and accurate feedback of potential risks during movement.

[0103] Step 3: Input the joint feature vector sequence into the variational autoencoder model for feature compression to obtain a low-dimensional latent variable representation sequence. Input the latent variable representation sequence into the time series prediction model and use a gated recurrent unit for time series modeling. Predict the latent variable representation of the next sampling point based on the latent variable representation of the previous sampling point in the latent variable representation sequence.

[0104] In this embodiment, the joint feature vector sequence is:

[0105]

[0106] in, For the first The joint feature vector of each sampling point and They represent the first Joint angle data and joint load data at each sampling point Indicates the first In the sampling points, the first Electromyographic signal data of a related muscle surface, This indicates the total number of muscles associated with the target joint. Indicates the total number of sampling points;

[0107] In this embodiment, the method for obtaining the low-dimensional latent variable representation sequence is as follows:

[0108] The joint feature vector of each sampling point is input into the trained variational autoencoder, which outputs the mean vector and log-variance vector of the latent variables. The latent variable representation for each sampling point is obtained by sampling from the latent variable distribution using a reparameterization technique, based on the following formula:

[0109]

[0110] in, and They represent the first The latent variable representation and mean vector of each sampling point , Represents the log-variance vector. Standard normal noise, i.e. It follows a multidimensional Gaussian distribution with a mean of 0 and a covariance matrix of identity.

[0111] A low-dimensional latent variable representation sequence is formed according to the timestamp order corresponding to the sampling points. The latent variable representation sequence is as follows: .

[0112] In this embodiment, in order to perform nonlinear compression on the acquired high-dimensional multi-source joint feature vector, extract the core representation features of the action behavior, and at the same time retain the joint expression relationship between joint angle, electromyographic signal and load information, the system adopts a variational autoencoder (VAE) model to extract low-dimensional latent variable representations. Through this model, the joint feature vector at each sampling point can be mapped into a low-dimensional latent variable vector with continuity and distribution characteristics, forming a latent variable representation sequence for subsequent time modeling and anomaly analysis.

[0113] Specifically, for each sampling point The collected joint feature vector Includes angle data of the target joint The RMS values ​​of electromyographic signals of multiple muscles associated with this joint. and the corresponding joint load data This vector serves as the input to the variational autoencoder model, which consists of an encoder network and a decoder network. The encoder is used to compress high-dimensional features into latent variable distribution parameters, and the decoder is used to reconstruct the original input for supervising model training.

[0114] The encoder network first receives the joint feature vector of the input. Through a fully connected neural network structure with several layers, it outputs two vectors: one is the mean vector of the latent variable distribution. One is the logarithmic variance vector. .

[0115] Since sampling directly from the probability distribution is non-differentiable and cannot be used for backpropagation training of neural networks, this embodiment employs reparameterization technology to achieve differentiable sampling. The reparameterization method transforms the random sampling process into a differentiable transformation between the model's learnable parameters and independent noise variables, specifically:

[0116]

[0117] in, This represents an independent Gaussian noise vector sampled from a multidimensional standard normal distribution. Through the above transformation, the latent variable representation corresponding to each sampling point is obtained. This latent variable can be optimized during training through gradient backpropagation.

[0118] The decoder network structure is symmetrical to the encoder, and its input is a latent variable. The output is the reconstructed joint feature vector. and with real input The reconstruction error is calculated. The training objective of the entire variational autoencoder is to minimize two loss functions: one is the reconstruction loss, which is used to ensure that the latent variables have fidelity and can recover the original multi-channel features; the other is the KL divergence loss, which is used to approximate the distribution of the latent variables to a standard normal distribution, thereby obtaining a coding space with a good distribution structure and continuity.

[0119] The overall training objective function of the variational autoencoder is:

[0120]

[0121] in, For reconstruction error, the second term KL divergence is used to measure the distance between the latent variable distribution and the standard normal distribution. The model uses the Adam optimizer for parameter updates, and the training data is the joint feature vector sequence of all sampling points. The training process is completed through repeated iterations of batch input, forward propagation, loss calculation, and backpropagation.

[0122] It is a variational distribution, representing a given input Afterwards, latent variables The conditional probability distribution, The output, calculated by the encoder network, is the mean of the latent variables. Sum of logarithmic variance Used to construct the Gaussian distribution of the latent variables. It is the reconstruction error, representing the error caused by the model passing through the latent variables. Reconstructed data Compared with actual input data The difference between them is represented here by the squared Euclidean distance, which is the sum of the squared errors of each feature. The goal of the reconstruction error is to minimize this difference, such that the input... It can be effectively reconstructed.

[0123] Represents the latent variable The expectation, due to From The reconstruction error is obtained from the probability distribution of the samples. Therefore, by calculating the expectation of all latent variables, the average reconstruction error is obtained. This part reflects the model's reconstruction ability. The reconstruction error measures the input... After being mapped to the latent space, can the decoder successfully convert the latent variables? Convert back to input data The quality.

[0124] The prior distribution of the latent variables is a standard normal distribution, i.e., a Gaussian distribution with a mean of 0 and a variance of 1. The prior distribution reflects the state of the variables before the input data is observed. At that time, latent variables The distribution that should be followed It is the KL divergence, used to measure the variational distribution. With prior distribution The smaller the KL divergence, the closer the variational distribution is to the prior distribution. The goal of KL divergence is to make the variational distribution as close as possible to the standard normal distribution, thereby ensuring that the latent space has a good distribution structure and avoiding unreasonable distributions of latent variables during training.

[0125] Finally, the trained variational autoencoder model can be used to infer the joint feature vector of new inputs. The system inputs the joint feature vector of each time point into the trained encoder, outputs the mean and variance of the corresponding latent variables, and obtains the low-dimensional latent variable representation of each sampling point based on reparameterized sampling. The latent variables of all sampling points form a latent variable representation sequence in chronological order, which will serve as the input basis for subsequent time series prediction models.

[0126] The method for predicting the latent variable representation of the next sampling point based on the latent variable representation of the previous sampling point in the sequence is as follows:

[0127] The trained gated recurrent unit is used to process the latent variable sequence. For time-series modeling, the latent variable representation of each sampling point is used as input. The initial hidden state of the gated loop unit is 0. It receives the latent variable representation of the current sampling point and the hidden state of the previous sampling point, and calculates the hidden state of the current sampling point through the gating mechanism.

[0128] Based on the hidden state of the current sampling point, a linear layer is used to predict the latent variable representation of the next sampling point. The underlying logic is as follows:

[0129]

[0130] in, Indicates the predicted first The latent variable representation predicted from each sampling point, Indicates the first The hidden state of each sampling point and These represent the learning weights and biases of the gated recurrent unit, respectively.

[0131] In this embodiment, to further model the dynamic evolution of individual joint motion behavior over time, the system represents the sequence of latent variables output by the variational autoencoder. Time series modeling is performed, and a latent variable trajectory prediction model is constructed using a gated recurrent unit (GRU). This model is used to capture the dependencies of latent variables in the time series and to predict the state of latent variables at future times, so as to assist in subsequent anomaly detection and risk analysis.

[0132] GRU is a recurrent neural network architecture with short-term memory capabilities. Its key feature is the selective memorization and forgetting of information through a gating mechanism, making it more efficient and stable when processing time series data. This model accepts each time step in the latent variable sequence. As input, and combined with historical hidden states, the updated hidden state at the current time step is output to pass information to the next time step. The first hidden state of the GRU network is a zero vector, ensuring that the model starts from an empty state to model behavior.

[0133] At each time step, the GRU network calculates intermediate candidate states and the final hidden state sequentially through its internal update and reset gate mechanisms, based on the current input latent variables and the hidden state of the previous time step. The update gate determines how much historical information should be retained in the current state, and the reset gate determines the degree of dependence between the current input and the historical state, thereby modeling the dynamic evolution trend of the latent variables. This hidden state can be regarded as an abstract encoding of the latent variable change trend of the previous time step. Based on the hidden state at the current moment, a set of linear transformation layers (i.e., fully connected layers) are used to predict and output the latent variable representation of the next time step.

[0134] During the training phase, to ensure the latent variable representations are effectively applied in subsequent anomaly detection and risk assessment steps, only healthy range-of-motion data are used during training. Specifically, multi-channel data such as joint angles, electromyographic signals, and joint loads are collected from the movement processes of healthy individuals, and these high-dimensional data are mapped to the latent variable space using a variational autoencoder (VAE). During training, only movement data within the normal physiological range of the human body are selected to ensure that the latent variable representations learned by the model conform to the physiological patterns of normal behavior. The selected training data is ensured to come only from healthy individuals, excluding any samples of pathological or abnormal movements. For example, for joint angle data, the system only collects angles within the normal range of motion, such as the range of motion during normal walking, running, and standing. This data is input into the VAE model as training data, processed by the encoder network to obtain latent variable representations, and captures the normal physiological patterns of movement amplitude, muscle activation patterns, and load changes in the latent variable space.

[0135] The encoder of a variational autoencoder (VAE) model maps the joint feature vector of each sampling point to a latent variable space, generating latent variables that represent the motion state of the target joint within its normal range of motion. To ensure that the distribution of the latent variable space conforms to the normal physiological range, the system can normalize the data by standardizing each feature of the input data to a specific interval, thereby normalizing the latent variables. The variational autoencoder training method in this embodiment ensures that the model learns only motion trajectories within the normal range of joint motion during training, which is of great significance for subsequent anomaly detection and risk warning. In this way, the system can effectively distinguish between normal and abnormal behavior and promptly identify abnormal transitions, thereby providing individuals with accurate risk assessment and early warning.

[0136] The GRU network structure mainly includes two core gating units: the update gate and the reset gate. The update gate controls how much memory from the previous time step is retained in the current hidden state, while the reset gate controls the degree of fusion between the current input and the past state. This mechanism can suppress the gradient vanishing problem when dealing with complex nonlinear time series, improve the model's ability to model long-term dependent behaviors, and thus improve the stability and accuracy of latent variable trajectory prediction.

[0137] The update gate controls the degree of updating of the hidden state at the current time step. Specifically, the update gate determines how much memory from the past should be retained in the current hidden state and to what extent it depends on the input information at the current time step. If the output of the update gate is large, it means that the information at the current time step should retain more of the hidden state from the previous time step, and vice versa, it means that it should depend more on the input data at the current time step. In this way, the update gate allows the network to flexibly adjust its dependence on past memories and current input during training.

[0138] The reset gate controls the degree of integration between the current input and the previous hidden state. The reset gate determines the degree of "forgetting" between the current input and the historical hidden state. When the output of the reset gate is small, it means that the current hidden state hardly depends on the historical state, but focuses more on the current input data. When the output of the reset gate is large, the historical hidden state has a stronger influence on the current hidden state. This mechanism enables GRU to flexibly adjust its memory of past information and its response to new information when facing non-stationary behavior or rapid changes in time series.

[0139] In this embodiment, the latent variable representation of each sampling point is a low-dimensional state expression compressed from multi-source data such as joint angle, electromyography signal and joint load at that moment. These state representations are highly time-dependent. In particular, during continuous movements, some posture changes are slow (such as the process of getting up), while some abnormal behaviors will cause abrupt changes in latent variables (such as falling down or a sudden increase in load). At this time, the gating mechanism will determine whether to retain the latent state of the previous moment to prevent the model from "misrecording" sudden abnormalities as normal trends.

[0140] After the data from the current sampling point is input into the GRU, the update gate first compares the current latent variable representation with the latent state information from the previous time step to determine whether the current prediction should rely more on past trends or should respond quickly to changes in the current data. For example, when an individual is standing or walking slowly, the update gate tends to retain more historical information to maintain continuity and prediction stability; while when an individual suddenly stands up from a squatting position and experiences a sudden increase in electromyography (EMG), the update gate will increase the weight of the current input, enabling the model to quickly perceive and respond to this physiological change.

[0141] The reset gate is primarily responsible for determining whether the current latent variable needs to be completely free from the influence of historical states, making it particularly suitable for dramatic posture transitions or aberrant leaps. Taking the knee joint as an example, when an individual suddenly jumps rapidly from a resting state, accompanied by drastic changes in electromyography and load, the reset gate reduces the influence of historical information on the current state, making the prediction of the current latent variable more sensitive to the initiation phase of the jump. Ultimately, through the synergistic effect of the update and reset gates, the latent state maintained by the GRU adaptively reflects the individual's current physiological state evolution trend at each sampling point. This mechanism has extremely high practical value for determining whether behavior evolves smoothly or whether aberrant leaps have occurred. It ensures that the model's predictions are not only a linear continuation of time but also a dynamic modeling of spatiotemporal changes in complex motion scenarios.

[0142] Step 4: Calculate the deviation between each actual latent variable and the predicted latent variable. Based on the deviation, determine whether there is abnormal transition behavior between the multi-channel data corresponding to adjacent sampling points. Map the abnormal transition behavior to a directed graph network, count the cumulative frequency and edge weight distribution of the abnormal transition behavior, classify the risk level, and output the corresponding musculoskeletal disease early warning signal.

[0143] In this embodiment, the deviation between the predicted latent variable representation and the actual latent variable representation for each sampling point is calculated using the following formula:

[0144]

[0145] in, Indicates the first The deviation between the predicted latent variable representation and the true latent variable representation for each sampling point. and They represent the first The predicted latent variable representation and the true latent variable for each sampling point, when When, determine the first The sampling point and the first There are abnormal transition behaviors between the multi-channel data of each sampling point. This indicates the deviation threshold.

[0146] In this embodiment, the latent variable representation sequence is extracted from the high-dimensional joint feature vector through a variational autoencoder model. It represents a low-dimensional compressed expression of joint angles, electromyographic signals, and load status. This latent variable space has the characteristics of continuity, predictability, and consistency with physiological behavior. Therefore, when the system predicts the state of latent variables at the next moment, the magnitude of its error actually reflects whether the current motion state conforms to the existing behavioral evolution law.

[0147] In the normal evolution of behavior, an individual's joint movements, muscle activation, and load transfer typically exhibit a continuous and gradually changing pattern. A GRU model trained on normal samples can accurately learn this time-series evolutionary trend, thus maintaining low bias when predicting the next state. If the current state is normal, the model will be able to "predict accurately"; conversely, if the current behavior undergoes a sudden change, this change will disrupt the original evolutionary trend, causing the model to be unable to accurately predict the next latent variable, and the error will increase accordingly. If the latent variable prediction bias exceeds a threshold, it indicates that the current behavior exceeds the evolutionary range that the model "considers normal," meaning that the behavior is unpredictable for the model. Since the model is trained based on healthy, continuous, and regular behavioral data, its "understanding" is itself built upon normal samples; therefore, its prediction error can naturally be used as an indicator of "abnormality."

[0148] Changes in latent variables represent the combined behavioral state of multi-channel physiological data (angle + electromyography + load) within a compressed space. Sudden changes in joint angle, intense muscle electrical activity, or load impacts can all cause "drastic shifts" in the latent variable space. These shifts are sensitively identified by the system's error judgment mechanism. This actually corresponds to "non-linear jumps in physiological behavior".

[0149] Regarding the setting of the deviation threshold, this embodiment provides two methods: the first is the static threshold setting method, which selects the 95th or 99th percentile as the threshold based on the prediction error distribution of normal samples in the training set, ensuring that the vast majority of normal behaviors will not trigger abnormal judgments; the second is the dynamic threshold adaptive method, which dynamically adjusts the threshold by combining the mean and standard deviation of the deviation within the individual's historical data or the sliding window. The numerical range enhances the individualized adaptability of anomaly detection. For specific use cases, such as rehabilitation training or industrial operations, the system can also use expert experience to set fixed thresholds to meet the accuracy requirements of specific behavior recognition.

[0150] In this embodiment, the method for outputting the corresponding musculoskeletal disease early warning signal is as follows:

[0151] Abnormal transition behaviors are mapped onto a directed graph network. The edge weights corresponding to each abnormal transition behavior are obtained, the frequency of abnormal transition behaviors is counted, and the transition risk state of the target joint during the observation process is calculated by combining the deviation corresponding to each abnormal transition behavior. The formula used is as follows:

[0152]

[0153] in, Indicates a transition risk status. This indicates the total number of sampling points. That is, the number of all edge weights. and They represent the first The edge weights and biases corresponding to each anomalous transition behavior. This represents the total number of all abnormal jump behaviors;

[0154] like This indicates a low-risk warning, suggesting occasional abnormal transitions and low edge weights. A medium-risk warning was issued, indicating frequent abnormal transitions. The system issues a high-risk warning indicating multiple abnormal jumps, occurring on high-weight paths. It issues an extremely high-risk warning that requires immediate intervention. In typical normal working and movement scenarios, occasional abnormal transitions (such as single electromyographic imbalances or slight abnormalities in joint angles) may be detected, but due to the low side weights and small biases, the overall risk score will remain at a low level. At this time, the RS value is usually between 0.05 and 0.15, corresponding to a low-risk level, and is only used for behavioral fluctuation alerts.

[0155] When an individual experiences multiple abnormal transitions within a short period, or when some transition paths have high structural sideweights (such as large-angle jumps, sudden increases in electromyography, or drastic load fluctuations), the RS value will gradually increase. Generally, in a mildly unstable state, the RS value is concentrated between 0.15 and 0.30, corresponding to a medium risk level, indicating a need to pay attention to potential abnormal patterns in behavioral evolution. In more severe cases, such as when an individual's behavior exhibits multiple consecutive high-intensity abnormal transitions, or when transitions are concentrated in areas already marked as high-sideweight paths by the system (such as deep bending, standing up under high load), the RS value will further increase to between 0.30 and 0.50. This range indicates that the system's predictive ability has failed multiple times, and the individual's behavior has significantly deviated from the normal trajectory, with the risk level identified as "high risk." If the RS value exceeds 0.50, it indicates the continuous detection of abnormal transitions with extremely high structural costs and drastic dynamic deviations, often accompanied by potential physiological damage risks or dysfunctional behaviors, such as violent falls, persistent misalignments, or severe muscle imbalances. At this point, RS has reached the "extremely high risk" threshold, which will immediately trigger an intervention-level warning mechanism, recommending that the current action be stopped and issuing a health risk warning.

[0156] In this embodiment, "abnormal transition behavior" is not equivalent to acute traumatic events in the traditional sense, but rather refers to potential physiological state abrupt changes caused by factors such as muscle fatigue, neural control imbalance, or decreased postural stability during continuous movement. This behavior is modeled using a combination of variational autoencoders and gated recurrent units to model multi-source time-series data, and a standard posture graph network structure is used to quantitatively analyze the dynamic evolution path, thereby identifying nonlinear anomalies in movement patterns. This modeling approach effectively captures the inherent characteristics of chronic accumulation in musculoskeletal disorders, making "abnormal transition behavior" an important indicator for predicting and warning of musculoskeletal disease risks, enabling early identification and intervention of potential injuries.

[0157] The risk status of a transition reflects two key dimensions of risk superposition: the structural burden of the transition itself and the degree to which the behavioral state deviates from normal healthy behavior. The edge weight reflects "how dangerous the path itself is." Each transition from one standard posture to another has a physiologically significant "load intensity," a value weighted by changes in angle abruptness, electromyography (EMG) bursts, and plantar weight-bearing. A large edge weight indicates a high physiological cost to the musculoskeletal system itself during the transition; even without abnormalities, it is considered a "high-cost movement." It is an inherent risk.

[0158] Latent variable bias reflects how drastically actual behavior deviates from the predicted direction of healthy behavior. It reflects the difference between the model's predicted normal healthy behavior and actual behavior, indicating a significant deviation in behavior and a situation prone to physiological damage. Combining both allows the model to consider both the physiological intensity of the action and the degree of prediction failure when assessing abnormal behavior, thus more comprehensively identifying potential risks.

[0159] A truly serious risk of musculoskeletal disorders exists only when both structural burden and behavioral state deviate from normal healthy behavior. If only the latent variable has a large deviation, but the edge weight of the current transition path itself is low, the risk cannot be said to be high. If only the edge weight is high, but the latent variable prediction is accurate, it means that although it is a high-cost action, the individual executes it very stably, and it is not considered high-risk. Only when both are combined does it constitute the most alarming situation in abnormal transitions.

[0160] Use normalization factor This is to standardize risk scores so that they are not affected by monitoring time or data length, which ensures that scores are comparable across different monitoring windows and assesses the "risk density" per unit time.

[0161] In this embodiment, the transition risk status scoring formula normalizes the cumulative risk contribution value of abnormal transitions, comprehensively considering the physiological burden of the transition action itself (reflected by the edge weights) and the difference between the model's predicted normal healthy behavior and the actual behavior (reflected by the latent variable bias). The product of the two constitutes a single score for transition risk. After statistically analyzing all abnormal transitions, the normalization process forms a comparable RS index, which can be used to drive the subsequent four-level risk warning mechanism.

[0162] In a typical monitoring cycle, assuming a sampling time window of 100 minutes and data collection every 30 seconds, the total number of transitions is approximately 200. Under normal circumstances, aberrant transitions typically account for a very small percentage of the total, generally not exceeding 10%. This means that out of 100 total transitions, at most 10 aberrant transitions may occur. The number of aberrant transitions is relatively small, and the weight of each aberrant transition is determined by differences in joint angle changes, electromyographic changes, and joint load. Based on physiological principles and movement patterns, the weight values ​​are generally between 0.3 and 1.5. Transitions with larger weights usually involve a wide range of joint movements or significant muscle activation. Latent variable prediction error reflects the gap between model predictions and actual observations, typically between 0 and 2. An error greater than 1 indicates a large prediction bias, but generally, latent variable errors are between 0 and 1. A bias exceeding 1 usually indicates very significant abnormal behavior.

[0163] Under normal circumstances, assuming no abnormal transitions occur (i.e., the summation term in the formula is 0), the transition risk state will be 0. This means that the system prediction is completely consistent with actual observations, with no abnormal events occurring, representing a "risk-free" state. However, assuming extreme abnormal behavior exists, which can lead to deviations in the maximum edge weight and maximum latent variable, the maximum value of the transition risk state also has a theoretical upper limit. Assuming a total of 100 transitions and 10 abnormal transitions, with a larger edge weight for each abnormal transition (let's say 1.5) and a larger prediction error for each latent variable (let's say 2), the calculated transition risk state is 1.5. Under common settings, the maximum value of the transition risk state is approximately 0.5. Of course, the probability of extreme abnormal behavior is very low, so such a high transition risk state rarely occurs in practice.

[0164] Under normal healthy conditions, the deviation between model predictions and actual behavior is small, and the frequency of anomalous transitions is also low. Therefore, the RS value usually ranges between 0 and 0.15. In other words, the system's behavior is stable under normal circumstances. When an individual's movement behavior experiences more frequent anomalous transitions, or when these transitions occur on paths with high joint loads and large angle changes, the RS value will gradually increase, possibly reaching 0.3 or higher. If the number of anomalous transitions is large and the prediction deviation is significant, the RS value will increase significantly, but it usually will not exceed 0.5. In extreme anomalous cases, the RS value can exceed 0.5.

[0165] Please see Figure 5 The present invention also provides a musculoskeletal disease risk assessment and early warning system based on multi-source data. The system is used to execute the above-described musculoskeletal disease risk assessment and early warning method based on multi-source data, comprising:

[0166] The feature acquisition module is used to continuously acquire multi-channel data of the target joint to form a joint feature vector sequence. The multi-channel data includes joint angle data corresponding to the timestamp, electromyographic signal data of the muscle surface associated with the target joint, and corresponding joint load data. The joint feature vector sequence includes joint angle sequence, electromyographic signal sequence, and joint load sequence.

[0167] The network construction module is used to perform K-means clustering on the joint angle sequence, divide all joint angle data into multiple standard poses, construct edge weights between any two standard poses based on the joint angle data, electromyographic signal data and corresponding joint load data within each standard pose, and construct a directed graph network representing the behavioral evolution path according to the collection order.

[0168] The variable prediction module is used to input the joint feature vector sequence into the variational autoencoder model for feature compression to obtain a low-dimensional latent variable representation sequence. The latent variable representation sequence is then input into the time series prediction model, and a gated recurrent unit is used for time series modeling. Based on the latent variable representation of the previous sampling point in the latent variable representation sequence, the latent variable representation of the next sampling point is predicted.

[0169] The risk assessment module is used to calculate the deviation between each actual latent variable and the predicted latent variable, determine whether there is abnormal transition behavior between the multi-channel data corresponding to adjacent sampling points based on the deviation, map the abnormal transition behavior to a directed graph network, count the cumulative frequency and edge weight distribution of abnormal transition behavior and classify the risk level, and output the corresponding musculoskeletal disease early warning signal.

[0170] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0171] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0172] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0173] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for musculoskeletal disorder risk assessment and early warning based on multi-source data, characterized in that, The specific steps include: Collecting multi-channel data of the target joint to form a joint feature vector sequence, the multi-channel data including joint angle data corresponding to a timestamp, electromyography signal data of a muscle surface associated with the target joint, and corresponding joint load data, the joint feature vector sequence including a joint angle sequence, an electromyography signal sequence, and a joint load sequence; Performing K-means clustering on the joint angle sequence to divide all joint angle data into multiple standard postures, and constructing an edge weight between any two standard postures according to corresponding joint angle data, electromyography signal data, and corresponding joint load data in each standard posture, and constructing a directed graph network representing an evolution path of a behavior according to a collection sequence; Inputting the joint feature vector sequence into a variational autoencoder model to perform feature compression and obtain a low-dimensional latent variable representation sequence, inputting the latent variable representation sequence into a time series prediction model, performing time series modeling by using a gated recurrent unit, and predicting a latent variable representation of a next sampling point based on a latent variable representation of a previous sampling point in the latent variable representation sequence; Calculating a deviation between each actual latent variable and a predicted latent variable, determining whether there is an abnormal transition behavior between multi-channel data corresponding to adjacent sampling points according to the deviation, mapping the abnormal transition behavior to the directed graph network, counting a cumulative frequency of the abnormal transition behavior and a distribution of the edge weight, and dividing a risk level to output a corresponding musculoskeletal disorder early warning signal; The logic of performing K-means clustering on the joint angle sequence is as follows: Determining a maximum activity angle of the target joint, dividing the maximum activity angle into angle ranges corresponding to K standard postures, and taking a median value of angles in each angle range as an initialized clustering center; Calculating an Euclidean distance between each joint angle data and each initialized clustering center, assigning the corresponding joint angle data to the nearest clustering center, and updating each clustering center to be a mean value of all angle values in the clustering cluster; Repeating the calculation of the Euclidean distance between each joint angle data and each updated clustering center, and reassigning the corresponding joint angle data to the nearest clustering center until the clustering centers converge, i.e., the clustering centers no longer change; Mapping the standard posture corresponding to the cluster where each joint angle data is located to the standard posture corresponding to each joint angle data.

2. The method of claim 1, wherein the method is based on multi-source data. The target joint is one of a hip joint, a knee joint, or an ankle joint in a lower limb joint, and the joint load data is an average pressure of both feet. When collecting the multi-channel data of the target joint, a plurality of sampling points are set, and joint angle data, electromyography signal data, and joint load data are synchronously collected at each sampling point, and the time intervals of adjacent sampling points are the same.

3. The method of claim 2, wherein the method is based on multi-source data. The method of constructing an edge weight between any two standard postures is as follows: Respectively calculating an average value of joint angles, a root mean square value of all electromyography signals, and an average value of joint loads in the internal clustering clusters of the two standard postures, and respectively calculating difference values of joint angles, electromyography signals, and joint loads between the two clustering clusters, and the formula is as follows: wherein, , and respectively represent the difference values of joint angles, electromyography signals and joint loads between the corresponding cluster groups when the standard postures are and , and respectively represent the average values of joint angles of the corresponding cluster groups when the standard postures are and , and respectively represent the average values of joint loads of the corresponding cluster groups when the standard postures are and , and respectively represent the root mean square values of electromyography signals of the corresponding cluster groups when the standard postures are and , , and both represent the numbers of the standard postures, and respectively represent the numbers of the target joint-associated muscles and the total number of the associated muscles. Weighting the difference values to construct an edge weight between the two standard postures, and the formula is as follows: wherein, represents a standard posture to the standard posture edge weight when migrating, , and respectively represent an angle weight, a signal weight, and a load weight, , and .

4. The musculoskeletal disorder risk assessment and early warning method based on multi-source data according to claim 3, characterized in that: The method of constructing a directed graph network representing an evolution path of a behavior is as follows: The joint angle data, the electromyographic signal data and the corresponding joint load data corresponding to each sampling point are taken as a time node, all time nodes are sorted according to the collection time stamps corresponding to the collection points, the direction of the directed graph network is from the previous time node to the next time node, and the edge weight of the adjacent two time nodes is the edge weight when the standard posture corresponding to the previous time node is migrated to the standard posture corresponding to the next time node.

5. The musculoskeletal disorder risk assessment and early warning method based on multi-source data according to claim 2, characterized in that: The joint feature vector sequence is: in, For the first The joint feature vector of each sampling point and They represent the first Joint angle data and joint load data at each sampling point Indicates the first In the sampling points, the first Electromyographic signal data of a related muscle surface, This indicates the total number of muscles associated with the target joint. Indicates the total number of sampling points; The method for obtaining the low-dimensional latent variable representation sequence is: The joint feature vector of each sampling point is input into the trained variational autoencoder, and the mean vector and the logarithmic variance vector of the latent variable are output, the latent variable representation corresponding to each sampling point is sampled from the latent variable distribution based on the reparameterization technique, and the formula is: wherein, and denote the latent variable representation and mean vector of the thsample point, respectively, , denotes the log-variance vector, is a standard normal noise, i.e. is drawn from a multi-dimensional Gaussian distribution with mean 0 and identity covariance matrix. The low-dimensional latent variable representation sequence is formed in chronological order of the time stamps corresponding to the sampling points, and the low-dimensional latent variable representation sequence is The method for predicting the latent variable representation of the next sampling point based on the latent variable representation of the previous sampling point in the latent variable representation sequence is: Using the trained gated recurrent unit on the latent variable sequence Temporal modeling is performed, the latent variable representation of each sampling point is taken as input, the initial hidden state of the gated recurrent unit is 0, the latent variable representation of the current sampling point and the hidden state of the previous sampling point are received, and the hidden state of the current sampling point is calculated through the gating mechanism. Based on the hidden state of the current sampling point, the latent variable representation of the next sampling point is predicted using a linear layer, and the logic is: wherein, represents a predicted latent variable representation for the th sample point prediction, represents a latent variable representation for the th sample point prediction, and respectively represent learned weights and biases for the gated recurrent unit.

6. The musculoskeletal disorder risk assessment and early warning method based on multi-source data according to claim 4, characterized in that: The deviation between the predicted latent variable representation and the actual latent variable representation of each sampling point is calculated, and the formula is: wherein, represents a deviation between a predicted latent representation of the th sample point and a true latent representation, and respectively represent a predicted latent representation of the th sample point and a true latent representation, when the deviation between the predicted latent representation of the th sample point and the predicted latent representation of the th sample point is greater than the deviation threshold, represents a deviation threshold.

7. The method of claim 2, wherein the method is based on multi-source data. The method for outputting the corresponding musculoskeletal disease early warning signal is: The abnormal transition behavior is mapped into the directed graph network, the edge weight corresponding to each abnormal transition behavior is obtained, the frequency of the abnormal transition behavior is counted, and the transition risk state of the target joint in the observation process is calculated in combination with the deviation corresponding to each abnormal transition behavior, and the formula is: wherein, denotes the transition risk state, denotes the number of all sampling points, denotes the number of all edge weights, and denote the edge weight and the bias corresponding to the th abnormal transition behavior, respectively, denotes the total number of all abnormal transition behaviors; If , a low-risk warning indicating an occasional abnormal transition with low edge weight is issued, if , a medium-risk warning indicating frequent abnormal transitions is issued, if , a high-risk warning indicating multiple abnormal transitions occurring on high edge weight paths is issued, if , a super-high-risk warning requiring immediate intervention is issued.

8. A musculoskeletal disorder risk assessment and early warning system based on multi-source data, characterized in that: The system is used for executing the musculoskeletal disease risk assessment and early warning method based on multi-source data according to any one of claims 1-7, and comprises: A feature collection module is configured to continuously collect multi-channel data of a target joint to form a joint feature vector sequence, the multi-channel data including timestamp corresponding joint angle data, electromyographic signal data of a muscle surface associated with the target joint, and corresponding joint load data, and the joint feature vector sequence including a joint angle sequence, an electromyographic signal sequence, and a joint load sequence. A network construction module is configured to perform K-means clustering on the joint angle sequence, divide all joint angle data into multiple standard postures, construct edge weights between any two standard postures according to corresponding joint angle data, electromyographic signal data, and corresponding joint load data inside each standard posture, and construct a directed graph network representing a behavior evolution path according to a collection sequence. A variable prediction module is configured to input the joint feature vector sequence into a variational autoencoder model for feature compression to obtain a low-dimensional latent variable representation sequence, input the latent variable representation sequence into a time series prediction model, perform time series modeling using a gated recurrent unit, and predict a latent variable representation of a next sampling point based on a latent variable representation of a previous sampling point in the latent variable representation sequence. A risk assessment module is configured to calculate a deviation between each actual latent variable and a predicted latent variable, determine whether there is an abnormal transition behavior between multi-channel data corresponding to adjacent sampling points according to the deviation, map the abnormal transition behavior into the directed graph network, count a cumulative frequency and an edge weight distribution of the abnormal transition behavior, divide a risk level, and output a corresponding musculoskeletal disease early warning signal.

Images