An individual motion behavior analysis system based on dynamic modal configuration and rule guidance fusion

Abstract

This invention discloses an individual motion behavior analysis system based on dynamic modal configuration and rule-guided fusion, belonging to the fields of intelligent perception, sports science, and human-computer interaction. It includes a data input module, a dynamic modal organization and configuration module, a multimodal feature modeling and fusion module, an indicator-level evaluation and behavior state prediction module, a statistical rule-guided analysis and result correction module, and a structured result generation and output module. These modules work together to form a complete closed loop for motion behavior analysis technology. This invention aims to address the problems in existing motion behavior analysis technologies, such as limited dimensionality of single-modal information, rigid multimodal fusion methods, insufficient stability and interpretability of analysis results, and strong dependence on large-scale manually labeled data.

Description

Technical Field

[0001] This invention relates to the fields of intelligent sensing, motion science and human-computer interaction, and specifically to an individual motion behavior analysis system based on the fusion of dynamic modal configuration and rule guidance. Background Technology

[0002] Individual motion behavior analysis is an important research direction in fields such as intelligent sensing, sports science, and human-computer interaction, and it has wide application value in scenarios such as sports training assessment, rehabilitation assistive devices, wearable systems, ergonomics analysis, and intelligent motion monitoring. By analyzing multi-source data generated during individual movement, it is possible to quantitatively describe movement status, motor coordination, and movement quality, providing data support for upper-level application systems or intelligent devices.

[0003] Currently, motion behavior analysis technologies mainly revolve around two technical paths: single-modal data processing and multimodal fusion modeling. However, significant engineering and methodological bottlenecks remain in practical applications. Single-modal data analysis methods generally suffer from limited information dimensions, sensitivity to noise and environmental conditions, and difficulty in comprehensively characterizing complex motion behaviors, resulting in insufficient stability and adaptability in complex or dynamic application scenarios. Existing multimodal fusion schemes often employ fixed modality combinations, typically based on feature concatenation or simple weighting methods; when some modal data is missing or of degraded quality, the overall analysis performance declines significantly. Furthermore, some multimodal analysis methods focus on improving model performance indicators, directly outputting prediction results, leading to insufficient interpretability and consistency of the results, making it difficult to meet the requirements of engineering systems for stable output and structured information representation. Summary of the Invention

[0004] The purpose of this invention is to provide an individual motion behavior analysis system based on the fusion of dynamic modal configuration and rule guidance. This system is based on multimodal fusion and rule guidance and is used to uniformly organize, fuse and analyze multi-source motion data and output structured information, thereby achieving flexible organization and collaborative processing of heterogeneous motion data such as inertia, posture, electromyography, and force.

[0005] To achieve the above objectives, this invention provides an individual movement behavior analysis system based on dynamic modal configuration and rule-guided fusion. The system includes: a multimodal data acquisition module for acquiring multi-source gait-related data of an individual during gait movement or rehabilitation training; a dynamic modal organization and configuration module for generating corresponding modal configuration parameters based on currently available modal types, thereby guiding subsequent feature modeling, model inference, and rule analysis processes. The modal configuration parameters indicate the participation status of each modal data point in the feature fusion and analysis process, ensuring consistent analysis logic and output structure under both single-modal and multimodal input conditions; a multimodal feature modeling and fusion module for performing feature modeling on each modal data point and generating a unified gait feature representation; and an index-level evaluation and behavior state prediction module for predicting behavior based on the fused gait data. The system represents gait characteristics and outputs multiple quantitative evaluation indicators and motion deviation state identifiers related to gait, providing accurate and interpretable quantitative basis for subsequent rule analysis. These multiple quantitative evaluation indicators and motion deviation state identifiers related to gait include at least gait spatiotemporal parameter indicators, symmetry and stability indicators, and multi-label prediction results related to motion deviation states. The rule-guided analysis and result correction module is used to perform rule constraints, logical verification, and consistency analysis on the indicator-level evaluation results based on recognized indicator definitions, statistical reference ranges, and behavioral evaluation logic in the field of gait biomechanics and motion behavior analysis, and to generate motion behavior state analysis results or structured evaluation information describing behavioral characteristics and their degree of deviation. The structured result generation module is used to uniformly organize the indicator-level evaluation results, rule analysis results, and statistical benchmark information to generate motion behavior analysis objects that meet the needs of engineering systems for structured data output and multi-scenario reuse.

[0006] Through the above technical solutions, the present invention has achieved the following beneficial effects: (1) By parallel modeling of multiple feature representation domains and framework-level multimodal fusion design, the organization and collaborative analysis of motion data with different physical attributes are realized. The implementation of the present invention can access multiple modal combinations including inertial measurement (IMU), surface electromyography (EMG), plantar pressure (Pressure), posture / joint capture (MoCap), etc. In engineering implementation, the system has exemplarily supported multiple modal combinations (exemplary configurations reach more than a dozen combinations), which significantly improves the adaptability and robustness under the conditions of sensor absence or modal combination changes compared with the traditional scheme that only supports a single modality or a fixed small number of modalities (exemplary experiments and implementation details are shown in the implementation examples and code records). (2) The introduction of a rule-guided and statistically distributed driven structured label generation mechanism reduces the dependence on manual annotation. In the cross-subject activity recognition experiment, under multiple independent training and testing conditions, GaitMF achieved an average accuracy of 62.21%±2.20% (mean±std) on the single-modal cross-subject task, and an exemplary accuracy of 88.57%±3.28% (mean±std) in the multimodal fusion experiment (stratified sampling). This shows that, without the need for large-scale manual annotation or under limited annotation conditions, the rule- and statistics-driven label / post-processing strategy can support engineering-usable recognition performance (see statistical results and code for specific experimental data and scripts). (3) Through the consensus-based adaptive fusion strategy and result correction mechanism, the output stability and consistency are effectively improved. In the cross-subject comparison experiment, the accuracy fluctuation of GaitMF (standard deviation ~2.20%) was significantly smaller than that of the baseline DeepConvLSTM (standard deviation ~4.48%), and the difference in accuracy was statistically significant in the paired t test. This indicates that the present invention exhibits stronger stability and consistency under different random initials, different individuals, and data perturbation conditions, thereby reducing the uncertainty of the "black box" output of a single model. (4) Using the index-level evaluation results as the core output object, the present invention can generate a structured analysis object containing multiple explicit biomechanical indicators such as cadence, stride length, symmetry, and stability (in practice, it can cover dozens of quantitative indicators), and the design ensures the traceability of key intermediate features, which facilitates the auditing and interpretability verification of the analysis process. This index-level output can be used as a standardized input for upper-level applications (such as rehabilitation training assessment, engineering alarms, and statistical analysis), enhancing the understandability and engineering adaptability of the system output.(5) The system adopts a highly intensive model design to balance performance and deployment feasibility: the exemplary deep learning model has 10,132 parameters (compared to the selected baseline model with 28,444 parameters, the parameter size is about one-third of the baseline model), which significantly reduces the model storage and transmission costs; single epoch measurement in the experimental environment shows that the training time is on the same order of magnitude under the same scale setting (see experimental record); combined with the JSON-based four-layer output structure (Meta / Metrics / Diagnosis / Analysis), the system can support fine analysis of single samples as well as meet the needs of statistical processing of batch samples and engineering system integration.

[0007] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0008] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a hierarchical diagram of the functional modules of the GaitMF multimodal motion behavior analysis system; Figure 2 This is a diagram of the overall system architecture of the present invention; Figure 3 This is a schematic diagram of the system functional hierarchy of the present invention; Figure 4 This is a comparison chart of signal preprocessing and filtering; Figure 5 This is a flowchart of the dynamic generation process of multimodal fusion weights driven by multi-source evaluation; Figure 6 This is a structure diagram of a multimodal, multi-task model; Figure 7 This is a multimodal performance comparison and evaluation chart; Figure 8 This is a graph showing the performance and robustness of multimodal fusion. Figure 9 This is a scatter plot of the accuracy of step frequency regression prediction; Figure 10 It is a confusion matrix heatmap; Figure 11 This is a diagram of the step frequency detection process based on autocorrelation; Figure 12 This is a graph showing the baseline distribution pattern of cadence in a population. Figure 13 It is a map showing the location of an individual within the population distribution; Figure 14 This is a schematic diagram of the decision-making process of the analytics engine; Figure 15 This is a diagram illustrating the chain of evidence. Figure 16 It is a statistical chart of the distribution of behavioral deviation triggers; Figure 17 It is a visualization of the decision boundary; Figure 18 This is a comparison of gait parameters between GaitMF and Imu_gait_analysis. Figure 19 Here is a detailed comparison table of GaitMF vs DeepConvLSTM; Figure 20This is a comparison of GaitMF and DeepConvLSTM activity recognition models; Figure 21 A comparison of the multimodal fusion capabilities of GaitMF and DeepConvLSTM; Figure 22 This is a visualization diagram of the batch structured analysis results of dynamic modal configuration; Figure 23 This is a schematic diagram of the system's single-sample analysis results. Detailed explanations of each diagram can be found in the subsequent implementation section. Detailed Implementation

[0009] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0010] (one) Figure 1 This is the overall technical solution of the present invention, namely, the functional module hierarchy diagram of the GaitMF multimodal motion behavior analysis system. The present invention is a structured individual motion behavior analysis and information generation system based on multimodal fusion and rule-guided methods, used for unified organization, fusion analysis, and structured information output of multi-source motion data. For example... Figure 1 As shown, the system of the present invention includes: (1) a multimodal data acquisition module; (2) a dynamic modality organization and configuration module; (3) a multimodal feature modeling and fusion module; (4) an index-level evaluation and behavior state prediction module; (5) a rule-guided analysis and result correction module; and (6) a structured result generation and output module.

[0011] The above modules operate collaboratively according to a preset process. The dynamic modal organization and configuration module improves the problem of "poor modal adaptability" in the existing technology by using dynamic weight normalization and modal validity detection to achieve flexible adaptation of different modal combinations. The multimodal feature modeling and fusion module and the index-level evaluation module are used to generate intermediate analysis indicators. Their output results serve as the quantitative input basis for the rule-guided analysis module. The system's final output is a structured motion behavior analysis result or information description object.

[0012] To aid in understanding the functional division and data flow relationships of the system at the overall level, a schematic diagram of the overall system architecture can be constructed, the structure of which is as follows: Figure 2 As shown. Figure 2 As shown in the diagram, this schematic illustrates the overall organization of the invention in terms of multimodal data input, feature analysis, rule reasoning, and structured result output from the perspective of system function and data processing flow. This diagram is only used to assist in understanding the functional composition and information flow of the system and does not constitute a limitation on the number of modules, connection methods, or specific implementation forms.

[0013] Based on the aforementioned overall system architecture, to further illustrate the functional layering and division of responsibilities at each layer in the engineering deployment and system implementation aspects of this invention, in another exemplary embodiment, a schematic diagram of a system functional layering architecture based on multimodal motion behavior analysis can be constructed, the structure of which is as follows: Figure 3 As shown.

[0014] like Figure 3 As shown in the diagram, this functional layering diagram logically divides the system of the present invention into layers from an engineering implementation perspective, exemplarily including a data layer, a model layer, a business layer, an application layer, and an access layer.

[0015] The data layer is used to uniformly manage multimodal motion data and intermediate processing results; the model layer is used to complete multimodal feature modeling, fusion analysis and index calculation; the business layer is used to perform rule reasoning, result consistency correction and information generation; the application layer is used to support business applications such as batch analysis, real-time analysis and visualization; and the access layer is used to provide command line interface, API interface or interactive analysis environment.

[0016] (II) Multimodal Data Acquisition Module The multimodal data acquisition module is used to acquire multi-source gait-related data of an individual during gait movement or rehabilitation training. The multimodal data is typically time-series continuous data, and its core characteristics and an example of an optional acquisition method are as follows: (1) Inertial Measurement Unit (IMU): contains multi-axis acceleration and angular velocity timing information, used to reflect the inertial dynamic characteristics during limb movement. An exemplary acquisition method is to deploy several inertial sensors on the subject's lower limbs and trunk to synchronously acquire motion signals during walking.

[0017] (2) Surface electromyography (EMG): This is a bioelectrical signal generated during muscle contraction, which can be used to characterize muscle activation state and synergistic relationships. An exemplary acquisition method is to place surface electrodes at key muscle groups in the lower limbs to acquire EMG signals during walking.

[0018] (3) Posture or joint position information (MoCap): used to describe the motion trajectory and posture changes of human joints in space. An exemplary acquisition method is to continuously sample key points of the joint based on optical or vision systems.

[0019] (4) Foot pressure or force distribution data (Pressure): used to reflect the weight-bearing situation of different areas of the foot during standing and walking. An exemplary acquisition method is to obtain the pressure distribution sequence through a foot pressure sensor.

[0020] The data mentioned above can be obtained from actual sensor acquisition, public datasets, or simulation-generated data, and is organized using a unified time base to ensure compatibility with subsequent processing modules. Before entering the analysis process, each modal data is resampled and time-aligned based on a timestamp, and segmented using a uniform window length to ensure that different modal features are fused and analyzed under the same temporal semantics.

[0021] To clarify the specifications and adaptation methods of different modal data in the system, in an exemplary embodiment, the modal dimension and encoder type can be configured, as shown in Table 1. It should be noted that this configuration is for illustrative purposes only and does not constitute a limitation on the technical solution of this invention.

[0022] Table 1 Modal Dimension Configuration Table To improve the reliability of input data, this module introduces a data quality scoring mechanism to evaluate the validity of the original data. In an example experiment, the accuracy rate of low-quality data screening can reach over 95%, effectively reducing the interference of noise on indicators such as step frequency and symmetry (Experiment 1: symmetry difference is only 5.1%). The calculation method is shown in the example of formula (17): in, The proportion of NaN values. The proportion of Inf values, The effective variance channel proportion (standard deviation) (Channel percentage). This scoring mechanism allows for the filtering of low-quality data before it enters subsequent analysis processes, thereby improving the stability of the overall analysis results.

[0023] To aid in understanding the improved usability of raw gait signals before they enter subsequent analysis processes, the changes before and after signal preprocessing can be visualized, as illustrated below. Figure 4 The comparison chart of signal preprocessing and filtering is shown.

[0024] Figure 4 Using inertial measurement signals as an example, this illustration demonstrates the changes in noise suppression and periodicity retention before and after preprocessing, aiding in understanding the supporting role of improved data quality in subsequent feature extraction and index calculation. This illustration is merely illustrative and does not constitute an essential component of the technical solution of this invention.

[0025] (III) Dynamic Modal Organization and Configuration Module The dynamic modality organization and configuration module is used to generate corresponding modality configuration parameters based on the currently available modality types, thereby guiding the subsequent feature modeling, model reasoning and rule analysis processes, so that the system can maintain stable analysis performance and output consistency even under conditions of modality missing or modality combination changes.

[0026] Through this mechanism, the system can automatically adjust the fusion strategy and analysis path according to the available modes. Compared with the traditional fixed mode combination or static weight scheme, it can effectively reduce the impact of mode loss on the overall analysis performance and improve the robustness and engineering adaptability of the system in complex application environments.

[0027] The modality configuration parameters are used to indicate the participation status of each modality data in the feature fusion and analysis process, so that the system can maintain consistent analysis logic and output structure under single-modal or multi-modal input conditions.

[0028] (1) Dynamic weight normalization When the modal combination changes, the initial weights of the effective modes need to be dynamically normalized to ensure that the sum of the weights is 1, as shown in the following formula: in, These are the normalized modal weights. The initial weights for the modes are shown in Table 2. This is the set of currently available modes. To prevent division by zero of small constants; If only IMU and Pressure data are obtained (EMG and MoCap data are missing), the normalized weights are: , Ensure that the total weight is 1.

[0029] (2) Fusion weight calculation function Figure 5 This invention demonstrates the dynamic generation logic of the fusion weights. Through evaluation of three dimensions—data quality, model consistency, and rule matching degree—it outputs fusion weights that reflect the credibility of the current modality, providing a quantitative basis for multimodal feature fusion.

[0030] The multimodal fusion method of this invention is not a static fusion method based on a single weighting formula. The core steps of the fusion process include parallel modeling of multimodal features and modality reliability assessment. The generated fusion weights need to be corrected by subsequent rule constraints.

[0031] The fusion weights are not fixed parameters obtained from neural network training, nor are they constants set by human experience; rather, they are intermediate variables generated during the model inference phase. These weights are determined by multiple sources of information, including data quality assessment results, model output consistency analysis results, and rule matching degrees constructed based on the statistical distribution of the normal population. The fusion weights do not participate in backpropagation training; their values ​​dynamically change with the state of the input data, reflecting the relative credibility of each modality under current conditions. This improves the system's stability and interpretability in the face of noise interference, modality loss, and cross-scenario applications.

[0032] The final output needs to be further validated for consistency through the rule-guided module to avoid a single model or modality dominating the decision-making process.

[0033] To clarify the logic for generating fusion weights for effective modes, this module uses a function... Taking into account data quality, model consistency, and rule matching degree, dynamic fusion weights for each modality are generated, as shown in the following formula: The input parameters (all normalized to [0,1]) are as follows: Modality The data quality is calculated using the data quality scoring formula in the "Multimodal Data Acquisition Module"; Modality Model consistency, namely the Agreement in the "multimodal feature fusion module"; Modality The rule matching degree is calculated as the number of rules matched for that modality. The total number of rules associated with this modality.

[0034] And the weighting coefficient: (Prioritize modal weights that ensure reliable data quality). (Second priority ensures stable modal weights for model output). (Ensure that the modal output conforms to the rule logic); satisfy This ensures the reasonable allocation of weights.

[0035] (3) List of key hyperparameters To illustrate the implementation of the multimodal fusion and dynamic configuration mechanism, the following parameters can be configured. These parameters are used to describe the relative importance of different modalities in the fusion process and the adjustment method; their specific values ​​can be adjusted according to the application scenario, data quality, or statistical results.

[0036] Table 2 List of Key Hyperparameters The initial weights are not fixed empirical constants, but rather an exemplary configuration based on the relative information contribution of different modalities in gait spatiotemporal structure, spatial posture, and force information. In other embodiments, they can be automatically determined through statistical analysis or learning methods. In this exemplary embodiment, the EMG modality is mainly used for muscle activation state analysis and anomaly assistance judgment, and is not included in the weight calculation of symmetry scoring. However, in other embodiments, the EMG modality can be incorporated into the fusion process of corresponding indicators as needed.

[0037] (4) Modal weight configuration table (used for asymmetry weighted calculation) Table 3 Modal Weight Configuration Table The weights shown in Table 3 are used for the weighted calculation of asymmetry in symmetry-related indicators. Since the symmetry indicators mainly rely on the comparison of spatial movement and force distribution of the left and right limbs, the above weights are used for the weighted calculation.

[0038] (iv) Multimodal Feature Modeling and Fusion Module The multimodal feature modeling and fusion module is used to perform feature modeling on each modality of data separately and generate a unified gait feature representation. To address the differences in physical properties, dimensional structure, and temporal characteristics among different modalities, a corresponding feature encoder can be configured for each modality, and the multimodal features can be integrated through the fusion unit.

[0039] (1) Encoder configuration In one exemplary embodiment, the parameter configurations of each modal encoder are shown in Table 4: Table 4. Encoder Detailed Configuration Table To illustrate the fusion logic of multimodal features, Figure 6 This schematically illustrates the structural form of a multimodal, multi-task deep learning model.

[0040] like Figure 6 As shown, this model configures dedicated feature encoding units for different modalities and generates a unified feature representation through a fusion mechanism, ultimately outputting multiple gait quantification indicators and motion deviation state prediction results. This model structure is only used to illustrate a feasible implementation method and does not constitute a limitation on network structure, number of layers, or parameter settings.

[0041] (2) Integration Strategy To achieve the complementary advantages of traditional methods and deep learning models, this module adopts a dynamic fusion strategy. Default weights are first set, and then adaptively adjusted based on data quality and model consistency, as detailed below: Default fusion weights: Adaptive weight adjustment: in, For data quality factors, As a model consistency factor, the clip function is used to limit the weight range and avoid extreme values ​​from affecting the fusion effect. This means that the default fusion weights are jointly averaged with the data quality factor and the model consistency factor to ensure that the fusion weights regress to the default strategy in low-confidence scenarios.

[0042] Intelligent fusion strategy: Weights are dynamically switched based on model consensus (Agreement). Consistency is calculated using the following formula, where i represents the sample index: For key indicators such as cadence, symmetry score, and stability score, after completing the consistency evaluation between traditional analysis results and deep learning prediction results, the fusion weights of the two types of results can be adaptively adjusted according to the consistency level. The exemplary rules are shown in Table 5.

[0043] Table 5 Intelligent Fusion Weight Table To evaluate the overall effectiveness of the multimodal fusion strategy under different modal combinations, the analytical performance of the system under various modal combinations was compared and analyzed, and the results are illustrated in the figure below. Figure 7 As shown.

[0044] The performance evaluation is based on statistical analysis of gait data from multiple subjects. The system performs analysis and reasoning using different modal combinations under the same sample set, and summarizes and statistically analyzes the results under each configuration condition.

[0045] like Figure 7 As shown, multimodal fusion outperforms single-modal configuration in overall performance, indicating that different modes have good complementarity in terms of gait temporal characteristics, spatial posture information, and force distribution characteristics. Among them, the combination of two or more modes shows more consistent performance in terms of accuracy and stability, while the performance of single-modal configuration fluctuates relatively more, indicating that multimodal fusion helps to reduce the risk of analysis instability caused by the failure of a single sensor or noise interference.

[0046] Furthermore, the risk level describes the degree to which motor behavior deviates from the statistical reference range, serving as an engineering prompt or behavioral quality assessment, but does not constitute a medical risk assessment. The predicted distribution of different risk levels is basically consistent with the overall distribution trend of the sample, verifying the rationality of the calibration of this invention in multi-category analysis scenarios. The performance indicators are calculated using a weighted statistical method to adapt to practical application scenarios with uneven category distributions.

[0047] (3) Model structure The model's multi-task output head simultaneously encompasses 47 gait quantification index regressions and 12 categories of motion deviation state classifications, providing comprehensive quantitative evidence for subsequent rule analysis and reflecting the core design principle of this invention: "feature modeling without directly outputting analytical conclusions." To further illustrate the engineering adaptability of this invention under conditions of modal missingness or data perturbation, in an exemplary embodiment, the system robustness is analyzed, and the results are as follows: Figure 8 As shown.

[0048] Figure 8 The system performance changes under different modal absence or noise interference conditions are illustrated to demonstrate the robustness of the invention in dynamic modal configuration scenarios. The results show that even with partial modal absence, the system maintains relatively stable evaluation performance, demonstrating the engineering adaptability of the framework-level design.

[0049] (v) Indicator-level assessment and behavior state prediction module The indicator-level evaluation and behavior state prediction module, based on the fused gait feature representation, outputs multiple quantitative evaluation indicators related to gait and motion deviation state identifiers, providing accurate and interpretable quantitative basis for subsequent rule analysis. The output results include at least gait spatiotemporal parameter indicators, symmetry and stability indicators, and multi-label prediction results related to motion deviation states, and do not directly constitute the final analysis conclusion, but only serve as the input basis for rule reasoning and structured analysis generation.

[0050] To comprehensively and accurately assess gait status, this module designs a quantitative index system from three core dimensions: symmetry, stability / variability, and gait frequency. The calculation logic and parameters of each index are explained below. The calculation methods described below are only used to illustrate one implementation of the present invention in this module, and their specific calculation forms do not constitute a limitation on the technical solution of the present invention.

[0051] (1) Symmetry index Gait symmetry is a core indicator for analyzing limb motor coordination, directly reflecting the muscle strength balance and neural control synchronicity of the left and right limbs. This module achieves precise quantification of symmetry through a process of "asymmetry calculation → multimodal weighted fusion → normalized scoring." 1) Asymmetry calculation: For different modalities, a unified formula is used to calculate the degree of asymmetry between the left and right sides, as follows: in, To prevent small constants from being divided by zero, and Application scenarios: IMU asymmetry (V=RMS acceleration / angular velocity signal), EMG asymmetry (V=mean (EMG power)), pressure asymmetry (V=∑pressure signal); 2) Overall symmetry score: To integrate the symmetry information of multiple modalities, a weighted average method is used to fuse the asymmetry of each modality, and then the result is normalized and mapped to a score in the [0,1] interval (the higher the score, the better the symmetry), as shown in the following formula: in, That is, the weighted average of the modal asymmetry ( For modal weights, refer to Table 2 for configuration.

[0052] (2) Stability / Variability Index Gait stability is directly related to motion safety risk assessment and motor control ability, while variability reflects the consistency of gait rhythm. This module employs a more robust calculation method to avoid outlier interference. Specific indicators are as follows: 1) Absolute deviation of the median: It is used to describe the dispersion of data and has strong robustness to outliers. The formula is as follows: in, A time-series data sequence with a certain step-state characteristic.

[0053] 2) Robust coefficient of variation: The coefficient of variation is calculated based on MAD (Magnitude of Variation) to replace the traditional standard deviation, further improving the tolerance for outliers. The formula is as follows: 3) Window variability: To address the dynamic characteristics of time-series data, the data is divided into multiple windows, and the degree of feature variation between windows is calculated to reflect the stability of gait rhythm. Window mean calculation: The default N = 4, and the window size = 128 / 4 = 32 frames.

[0054] Quantification of variability: 4) Overall stability score: After weighted fusion of multimodal variability, it is transformed into a score in the [0,1] interval through nonlinear mapping (the higher the score, the better the stability), as shown in the following formula: in, The weighted average of modal variability is calculated.

[0055] (3) Step frequency estimation formula Step frequency is a core indicator of gait spatiotemporal parameters, reflecting the speed of gait rhythm. This module achieves accurate step frequency detection based on autocorrelation analysis, and the process is as follows: 1) Signal preprocessing The IMU data is mean-centered to remove baseline drift, using the following formula: in, For IMU time series data, This is the global mean of the IMU data.

[0056] 2) Normalized autocorrelation: The autocorrelation function of the signal is calculated to capture the periodicity of the gait cycle, as shown in the following formula: in, For the lagging step size, Frame (corresponding to approximately 180 steps / minute) The frame (corresponding to approximately 40 steps / minute) covers the step frequency of gaits that fall within the statistical reference range.

[0057] 3) Period detection and step frequency calculation Find the first peak of the autocorrelation function within the effective lag range [10, 60], which corresponds to the gait period. The formula is as follows: Where τ* represents the number of lag frames corresponding to the gait cycle, in units of frames.

[0058] The step frequency (unit: steps / minute) is calculated based on the period, where FPS=30 (example sampling frame rate), and the formula is as follows: The effective range of cadence is [40, 180] steps / minute. If it exceeds this range, it is judged as an abnormal gait rhythm.

[0059] To verify the predictive consistency of the index-level regression task, the model's prediction results can be compared with the reference values. The results are illustrated in the figure below. Figure 9 As shown.

[0060] Figure 9 This figure illustrates the correspondence between the model's predicted gait frequency and the reference value, used to verify the predictive accuracy of the gait frequency regression task. The scatter distribution being close to the diagonal indicates that the model has good predictive consistency within the statistical reference gait rhythm range; this figure is only used to illustrate the model's performance at the indicator-level prediction level.

[0061] To evaluate the overall performance of the motion deviation state classification task, the prediction results can be visualized using a confusion matrix, as illustrated below. Figure 10 As shown.

[0062] Figure 10 The distribution of prediction results for the motion deviation state classification task is presented as a confusion matrix, used to evaluate the model's ability to identify different anomaly categories. This figure primarily illustrates the model's overall performance in distinguishing motion deviation states and is not intended as a direct basis for the final analytical conclusions.

[0063] To aid in understanding the calculation logic of step frequency estimation in this module, in an exemplary embodiment, a processing flow based on autocorrelation analysis can be adopted, as illustrated below. Figure 11 As shown.

[0064] The upper part shows the input signal after mean centering, and the lower part shows the normalized autocorrelation function curve and the physiological gait search window. This diagram schematically illustrates the preprocessed input signal and its normalized autocorrelation function curve, where the first peak of the autocorrelation function corresponds to the gait cycle position and is used to calculate the gait frequency index. This process diagram is used to illustrate the basic idea of ​​gait frequency estimation and does not constitute a limitation on the specific implementation process or parameter settings.

[0065] (4) Core Regression Task List To clarify the module's output indicator system and facilitate subsequent rule matching and result verification, the core quantitative indicators are summarized in the table below, covering the evaluation parameters and value ranges of key gait dimensions: Table 6. List of Core Regression Tasks The aforementioned indicators cover the spatiotemporal characteristics, coordination ability, stability, and data reliability of gait, providing comprehensive and standardized quantitative inputs for rule-guided analysis and evaluation. This ensures the accuracy and interpretability of the generated gait quality analysis results, structured indicator sets, and other information, and can serve as standardized data support for subsequent use by platforms, fitness equipment, or rehabilitation aids.

[0066] Statistical results based on the exemplary reference population data Figure 12 The distribution pattern of the step frequency index is shown. The purple histogram reflects the frequency distribution of the sample, the curve is the result of probability density fitting, the dashed line marks the population mean, and the light purple area is the range of ±1 standard deviation.

[0067] This distribution exhibits good near-normal characteristics, demonstrating the statistical distribution characteristics of the cadence index in the reference population sample, and illustrating the feasibility of establishing a relative evaluation benchmark through population distribution. This distribution result is for illustrative purposes only and does not limit specific data sources or statistical methods.

[0068] To aid in understanding the relative positional relationship of individual gait indicators within a statistical benchmark, in one exemplary embodiment, individual indicators can be mapped to population distribution intervals, as illustrated below. Figure 13 As shown.

[0069] like Figure 13 The illustration shows the relative position of an individual's gait indicators within the statistical distribution of the population, illustrating the basic idea behind this invention's risk assessment based on statistical benchmarks. By mapping individual indicators to population distribution intervals, the relative quantification of the degree of anomaly can be achieved, rather than relying on statistical thresholds.

[0070] (vi) Rule-guided analysis and result correction module The rule-guided analysis and result correction module is the core functional module of this invention for achieving consistency, stability, and interpretability of analysis results. Its main function is as follows: Based on the recognized indicator definitions, statistical reference ranges, and behavioral evaluation logic in the field of gait biomechanics and movement behavior analysis, the indicator-level evaluation results are subject to rule constraints, logical verification, and consistency analysis, thereby avoiding the problems of result fluctuations or inexplicability caused by relying solely on the output of a single model.

[0071] When the model analysis results deviate significantly from the rule constraints, this module uses rule correction and consistency adjustment mechanisms to modify or reduce the weight of the analysis results, thereby improving the analysis stability across different data conditions, modality combinations, and individuals. The output of this module is the analysis results of motor behavior states or structured assessment information, used to describe behavioral characteristics and their degree of deviation, and does not directly constitute a medical diagnosis or clinical decision.

[0072] (1) Rule definition and analysis logic The rule system adopted in this invention is not based on empirical threshold settings, but on a statistically constructed method. It is built on the basis of the recognized indicator system and statistical evaluation methods that have been formed in the fields of gait analysis, biomechanics and rehabilitation assessment over a long period of time. The source of its rules mainly includes the following three levels.

[0073] 1) Source of the authoritative definition of general gait indicators The core gait indices involved in this invention (including cadence, stride length, gait speed, symmetry, stability / variability, etc.) all adopt universally accepted and standardized definitions in the field of gait analysis. Their indices and measurement logic are derived from internationally recognized authoritative textbooks and review articles, such as: Perry&Burnfield: "Gait Analysis: Normal and Pathological Function" Winter: "Biomechanics and Motor Control of Human Gait" Kirtley: "Clinical Gait Analysis: Theory and Practice" The aforementioned teaching materials systematically provide the physical meaning, calculation methods, and behavioral interpretation logic of gait parameters, serving as a general reference standard in the fields of sports science, rehabilitation assessment, and clinical gait analysis. This invention strictly adheres to the aforementioned general definitions at the indicator level, without introducing custom or non-standard indicators.

[0074] 2) Basis for constructing the normal range and statistical thresholds of indicators To avoid insufficient generalization ability of rules due to the use of fixed statistical thresholds, the statistical thresholds of each indicator in this invention are derived from the gait statistical distribution characteristics (normative data) of healthy people and are calibrated in combination with the sample data of this invention.

[0075] The construction of rule thresholds is based on, but is not limited to: statistical research results of gait parameters of healthy adults (such as mean and standard deviation); statistical results of large sample populations based on inertial measurement units (IMU), motion capture (MoCap) and plantar pressure data; and reference ranges given in publicly available gait standard databases and review studies.

[0076] In its specific implementation, this invention employs one or a combination of the following statistical methods to determine the rule boundaries: a) Mean ± multiple standard deviations; b) Median and percentile range (e.g., 5%–95%); c) Robust statistical intervals based on distribution patterns.

[0077] The threshold constructed in this way has clear statistical significance, rather than being an artificially set constant, and can be adaptively adjusted according to the target population, device type, or data quality.

[0078] 3) Consistency between rule logic and behavioral assessment consensus In addition to numerical thresholds, the rule system of this invention also follows the general logic in the field of motor behavior assessment and rehabilitation analysis, namely: A deviation of a single indicator from the statistical reference range reflects an abnormal trend in the corresponding behavioral dimension; Joint deviations of multiple indicators within the same or cross-dimensional contexts reflect more significant behavioral abnormalities. The changing trends of indicators such as stability, variability, and symmetry are used to describe behavioral control ability and rhythm consistency.

[0079] For example: Symmetry indicators are used to reflect the coordination of left and right limb movements; Stability or variability indicators are used to reflect the consistency of movement rhythm and the degree of control stability; The plantar pressure distribution index is used to reflect the load-bearing balance characteristics; The cadence index is used to reflect the overall rhythm of movement.

[0080] The above rules and logic are consistent with existing gait assessment and rehabilitation analysis practices, and are only used for behavioral state analysis and information generation, without involving the direct output of medical diagnostic conclusions.

[0081] (2) Exemplary rule definition Based on the above rule sources and statistical methods, examples of behavior analysis rules can be constructed as shown in Table 7.

[0082] Table 7 Definition of Core Analysis Rules It should be noted that the above thresholds are merely illustrative examples, and the specific values ​​can be adjusted based on the statistical characteristics of the target population, the conditions of the data collection equipment, and the application scenario, without affecting the rule construction method and overall analysis process of this invention.

[0083] When data for a particular modality is missing or deemed invalid, the rule items associated with that modality will be automatically skipped or downgraded to avoid analytical bias caused by incomplete information.

[0084] (3) Result confidence assessment and rule consistency correction To further quantify the reliability of the structured analysis results, this module designs a multi-dimensional confidence assessment mechanism to label the credibility of the structured analysis results, specifically including: 1) Analyze confidence level: Taking into account three factors—data quality, model consistency, and rule consistency—the overall confidence level of the analysis results is calculated using the following formula: in, For data quality scoring (same as the formula in scheme (II)), Indicates the degree of consistency between different analysis paths; Indicates rule consistency (1.0 for no conflicting rules, decreasing when there are conflicts). Rule consistency is used to measure the relative deviation between results under different analysis paths or rule constraints.

[0085] 2) False label confidence (Self-Training scenario): in, Indicates sample consistency. Indicating rule matching degree, in one exemplary embodiment, the training sample selection threshold is... .

[0086] 3) Rule-guided analysis process and effect description To aid in understanding the overall logic of the rule-guided analysis module, in one exemplary implementation, a schematic diagram of the analysis engine's decision-making process can be constructed (e.g., Figure 14 As shown in the figure, this diagram fully illustrates the key steps of the entire process, starting from the input of multi-source sample data, sequentially covering the dual-path feature extraction of "traditional analysis module + multi-task deep learning model", the screening of analysis-related features, the adaptive fusion strategy based on consistency, as well as the feature mapping form of the analysis rule engine, the judgment of indicator anomalies, the matching of risk levels, and finally the multi-format output of the analysis results.

[0087] Furthermore, a schematic diagram of the chain of evidence for analysis can be constructed (such as...). Figure 15 As shown in the figure): Taking sample #18 as an example, the output results of "deep learning analysis" and "traditional analysis" in terms of core indicators such as step frequency, symmetry, and stability are compared. At the same time, the comprehensive results generated by the adaptive fusion strategy are presented, which clearly reflects the complementarity of multi-source analysis results and the formation logic of the final conclusion.

[0088] Perform statistical analysis on the trigger distribution of rules at the sample set level (e.g.) Figure 16 As shown in the figure, it can be used to evaluate the rule coverage and distinguishing ability.

[0089] Furthermore, the analysis boundaries of different state intervals are visualized using a low-dimensional projection method (e.g., Figure 17 As shown, this is used to help understand the distinguishing logic of the results of structured behavior analysis, and the presentation does not limit the specific boundary form.

[0090] (vii) Structured Result Generation Module The structured results generation module is used to uniformly organize indicator-level evaluation results, rule analysis results, and statistical benchmark information to generate standardized motion behavior analysis objects, so as to meet the engineering system's needs for structured data output and multi-scenario reuse.

[0091] (1) Example of comprehensive behavioral quality score To quantitatively assess the overall gait quality of an individual, this module selects symmetry score and stability score as core evaluation indicators to construct a comprehensive behavioral quality score, as shown in the following example: in, It characterizes the degree of coordination and balance between the left and right limbs during the gait cycle. This characterizes the overall stability and safety during the gait process. The weighting settings mentioned above reflect the relative importance of different indicators in behavioral state assessment, and their specific proportions can be adjusted according to the application scenario and data distribution characteristics.

[0092] The comprehensive gait quality score ranges from [0,1] and can be classified into risk levels based on the statistical distribution characteristics of historical samples and risk assessment experience. For example, a score ≥ 0.85 is considered a low-risk range, a score between 0.7 and 0.85 is considered a medium-risk range, and a score < 0.7 is considered a high-risk range.

[0093] It should be noted that the above weighting ratios and risk range divisions are for illustrative purposes only and do not constitute a limitation of the present invention. The weighting coefficients and risk thresholds can be adjusted according to actual needs in different application scenarios or for different target populations.

[0094] (2) Structured Behavior Analysis Objects The structured behavior analysis object adopts a standardized field design to ensure a consistent output format. It specifically includes the following four core fields, and the functions and application scenarios of each field are as follows: 1) Set of behavioral quantitative indicator fields: It includes numerical analysis results for multiple parameters such as step frequency, stride length, symmetry, and stability, presented in numerical form. 2) Behavioral characteristic deviation identifier field: Anomaly alerts triggered by annotations directly correspond to the analysis rules in the rule definition table, making anomalies immediately apparent. 3) Behavioral status level and confidence level fields: The overall behavioral state is described in a hierarchical manner, and corresponding confidence level labels are provided; 4) Behavioral analysis description field: Natural language analysis descriptions automatically generated based on rule matching results are used to aid in understanding the analysis conclusions.

[0095] The structured analysis object contains descriptive text constrained by rule matching results, indicator deviation patterns, and preset semantic templates. This text generation does not rely on an open generation model, thereby ensuring the consistency and interpretability of the output content and analysis results. The structured analysis object can be exported as a database record, CSV, JSON, or HTML report format for subsequent system calls, data analysis, or interactive display. It should be clearly stated that it does not constitute a medical diagnostic conclusion. The risk level described in this invention is only an engineering classification description of the deviation of exercise behavior from the statistical reference range and is not used for disease or health risk assessment.

[0096] From an experimental verification perspective, the technical rationality and effectiveness of the individual motion behavior analysis system based on multimodal fusion and rule guidance proposed in this invention are explained. Comparative experiments with existing representative methods on the same publicly available dataset verify the system's technical performance in terms of gait parameter calculation consistency, model analysis capabilities, and multimodal fusion effects. The above experimental results illustrate the feasibility and engineering effectiveness of the technical solution of this invention.

[0097] Explanation of Experimental Data and Comparison Principles To ensure the objectivity and reproducibility of the experimental results, all experiments were conducted based on the publicly available HuGaDB (HumanGait DataBase) real gait dataset. This dataset contains multimodal sensor data collected from 18 subjects under different motion states, with a sampling frequency of 60 Hz, covering various commonly used motion analysis modalities such as inertial measurement unit (IMU) signals and electromyography (EMG) signals. It is used as a publicly available reference dataset in gait analysis and human activity recognition research.

[0098] In the experimental design, this invention follows the following basic comparison principles: (1) The comparison methods are all representative methods commonly used in academia or engineering practice; (2) Comparative experiments were conducted under the same dataset, the same training / test partition, or the same input conditions; (3) The comparison focuses on technical rationality, consistency and integration gain, rather than the extreme optimization of a single performance index.

[0099] Based on the above principles, three sets of technical verification experiments were designed and completed, corresponding to the verification of the rationality of gait parameter algorithms, the verification of deep learning model analysis capabilities, and the verification of multimodal fusion capabilities.

[0100] (II) Experiment 1: Consistency verification of gait parameter calculation (compared with traditional methods) The comparison methods selected are representative and mature methods in the field (such as imu_gait_analysis and DeepConvLSTM), which are often used as baseline models or control methods in relevant review literature.

[0101] (1) Experimental objective This experiment aims to verify the rationality and consistency of the proposed system at the gait parameter calculation level. It should be clarified that the traditional imu_gait_analysis method is a "data processing tool," while GaitMF is an "end-to-end analysis system from data processing to analytical decision-making." This experiment only verifies the equivalence of their capabilities at the first layer (data processing). By comparing the key gait indicators output by GaitMF with the calculation results of traditional signal processing methods, this experiment evaluates whether the proposed system can obtain analytical results consistent with commonly used methods in academia without relying on manual parameter tuning.

[0102] (2) Comparison methods and experimental procedures The traditional `imu_gait_analysis` method was selected as the comparison object. This method extracts gait period and related parameters based on classic signal processing procedures (Butterworth low-pass filtering, gyroscope amplitude synthesis, peak detection, etc.), and belongs to the gait parameter calculation methods widely used in existing research and engineering practice. The specific experimental procedure includes: 1) Perform unified format conversion, resampling to 200Hz (consistent with the GaitMF standard to use a unified time resolution for subsequent windowed analysis) and physical quantity normalization on the raw HuGaDB data; 2) The same gait data segment was analyzed using both the traditional method and GaitMF (the traditional method is peak detection + period calculation, while GaitMF is window segmentation + autocorrelation period detection). 3) Statistical comparison of the core gait indicators such as step frequency and symmetry output by the two methods.

[0103] (3) Experimental results and analysis Under the condition of processing approximately 23 minutes of continuous gait data (281,070 data points) and a total of 4,363 analysis windows, the comparative results show that: 1) The two methods show consistent trends in core indicators such as step frequency and symmetry (the traditional method has a step frequency of 78.0 steps / min, while GaitMF has a step frequency of 69.6±5.3 steps / min, both of which are within the acceptable error range for engineering). 2) The average difference in symmetry indices was 5.1%, which is within the common range of algorithmic consistency errors in gait analysis and biomechanical engineering research. (Perry & Burnfield (2010) pointed out in authoritative research on gait analysis that due to differences in gait event detection methods, signal preprocessing strategies, feature definition methods, and sampling conditions, even for the same subject, different analysis methods or implementation frameworks usually have certain deviations in the calculation results of gait parameters such as symmetry and variability. In engineering practice and biomechanical analysis, for continuous parameters such as gait symmetry and gait variability, when the relative difference between different methods is within the range of about 5%–10%, it is generally considered to be within an acceptable range of consistency.) 3) GaitMF provides additional information not available in traditional methods, such as confidence level (±5.3), stability score (0.489), and window-level analysis results (561 valid windows).

[0104] Figure 18 The results show that the system of the present invention has good consistency with traditional academic methods in gait parameter calculation, which verifies its rationality as a basis for subsequent analysis.

[0105] (III) Experiment 2: Validation of deep learning model analysis capabilities (compared with DeepConvLSTM) (1) Experimental objective This experiment aims to verify the basic analytical capabilities of the deep learning analysis module in the system of this invention, evaluate whether it has behavior recognition and feature modeling capabilities comparable to classic deep learning models under single-modal (IMU) conditions, and verify the engineering advantages of its lightweight architecture.

[0106] (2) Comparison of model and experimental setup The DeepConvLSTM-Lite model was selected as the baseline for comparison (4×Conv1D (32,5) + LSTM (32), with 28,444 parameters). This model is one of the representative deep learning baseline models commonly used in the field of human activity recognition. The model architecture of this system is multi-scale CNN + BiLSTM (16) + Attention (10,132 parameters).

[0107] The experiment was conducted based on 36 columns of IMU modal data from the HuGaDB dataset, with the following core settings: 1) The input window length is 24 frames, and a cross-subject evaluation strategy is adopted (the training and testing subject sets do not overlap). 2) Training epochs: 5, batch size: 256, optimizer: Adam, loss function: CategoricalCrossentropy; 3) The task type is 12 categories of human activity recognition, with 13,142 samples in the training set and 9,337 samples in the test set.

[0108] (3) Experimental results and analysis This experiment was conducted under cross-subject evaluation conditions for a multi-class recognition task of 12 categories of human activities (in the multi-class scenario, the accuracy of random guessing was 1 / 12≈8.33%, which is not 50% of that of a binary classification task).

[0109] It should be noted that under cross-subject evaluation, short window input, and single-modal conditions, the overall recognition accuracy is usually significantly lower than the results under the same subject or multimodal settings. This experiment focuses on the relative performance differences of different models under the same conditions, rather than the absolute accuracy values.

[0110] Experimental results show that the accuracy of this model in this task is 52.58% (significantly higher than the random benchmark of 8.33%), and the weighted F1 score is 53.98%. Compared with the comparison model, the performance gap (accuracy gap of 2.10% and weighted F1 gap of 2.18%) is less than 3%, but the number of parameters of this model is only 35.6% of that of the comparison model (10,132 vs 28,444), a reduction of 64.4%, which significantly reduces the model complexity. The training time only increases by 9% (3.6s / epoch vs 3.3s / epoch), while also having additional advantages such as attention visualization interpretability and multimodal extension support.

[0111] Figure 19 and Figure 20 The results show that the system of the present invention still has analytical capabilities comparable to those of classic deep learning models under lightweight conditions, meeting the requirements of computational efficiency for engineering deployment and practical applications.

[0112] This experiment prioritizes the comparison of typical deep learning baseline models with similar structural size and engineering deployment complexity to ensure fairness in the comparison.

[0113] (iv) Experiment 3: Verification of multimodal fusion capability (comparison of fusion gain) (1) Experimental objective This experiment aims to verify the information fusion capability of the present invention under multimodal conditions, focusing on evaluating whether the system can effectively mine complementary information between IMU and EMG modes, thereby achieving a performance improvement superior to single-modal analysis, and verifying whether its fusion capability is superior to the classical baseline model.

[0114] (2) Experimental design and comparison methods The experiment was conducted using 20% ​​stratified sampling of the HuGaDB dataset (13,008 samples, 9756 in the training set and 3252 in the test set). The task was to identify 12 types of activities, and the core settings were as follows: 1) Single-modal input: IMU (36-dimensional), dual-modal input: IMU+EMG (38-dimensional); 2) Window size: 64 frames; Training epochs: 8. 3) Define "fusion gain = dual-modal performance - single-modal performance" and compare the fusion gain difference between this system and the DeepConvLSTM model.

[0115] (3) Experimental results and analysis Experimental results show that: Under single-mode conditions, the F1 value of this system is 86.98%, while that of DeepConvLSTM is 86.31%, and the performance of the two is similar. Introducing the EMG mode significantly improves the performance of both models, but the fusion gain of this system is significantly higher. 1) F1 score: This system improved by 2.75% (86.98% → 89.73%), while DeepConvLSTM improved by 1.96% (86.31% → 88.27%), resulting in an advantage of 0.79%. 2) Accuracy dimension: This system increased by 2.45% (89.06%→91.51%), DeepConvLSTM increased by 1.37% (89.42%→90.79%), with an advantage of 1.08%; Under dual-modal conditions, the absolute performance of this system (Accuracy 91.51%, F1 89.73%) is superior to the comparison model in all aspects.

[0116] Figure 21 The results show that the multi-scale encoder + attention fusion architecture adopted in this invention is more effective in utilizing multimodal complementary information.

[0117] (V) Comprehensive Analysis and Technical Conclusions The above three sets of experiments show that: (1) The present invention maintains good consistency with traditional academic methods in gait parameter calculation (symmetry difference is only 5.1%), which verifies the rationality of the data processing layer; (2) In single-modal analysis tasks, the system of the present invention achieves analysis capabilities comparable to the classic DeepConvLSTM model with only 35.6% of the parameters, with significant lightweight advantages, meeting the needs of engineering deployment; (3) Under multimodal conditions, the fusion gain of the system of the present invention is superior to the classical baseline model in all aspects, which reflects the technical advantages of the multi-scale encoder + attention fusion architecture; The above experimental results verify the rationality, effectiveness and engineering feasibility of the technical solution of the present invention from multiple perspectives, and provide experimental basis for its application in actual motion analysis and individualized assessment scenarios.

[0118] In recent years, Transformer-based models have achieved some performance improvements in activity recognition tasks, but their model size and computational complexity are high, typically relying on large training samples and computing resources. Considering the engineering deployment requirements of the system in this invention and the interpretability requirements of the rule-guided analysis framework, this type of model was not introduced for comparison in this stage of the experiment. Future work can introduce more complex models for extended verification while maintaining the overall system architecture.

[0119] For example The sample size, parameter configuration, and result data in the following embodiments are exemplary and used to illustrate the implementation and analysis effects of the present invention under different application conditions. They do not constitute a limitation on actual application effects or performance indicators. Those skilled in the art can adjust or replace the relevant parameters without departing from the overall concept of the present invention.

[0120] (I) Example 1: Structured Motion Behavior Analysis Process under Multimodal Conditions (1) Implementation conditions In one embodiment, several subjects are selected as example samples, and their multimodal gait data during walking are collected.

[0121] The multimodal data includes at least two or more of the following modalities: inertial measurement data, surface electromyography signals, joint motion or posture information, and plantar pressure distribution information.

[0122] The modal data are aligned and organized according to a unified time reference to form temporally consistent multimodal input data, which is then input into the individualized motion behavior analysis system described in this invention as input data for the analysis process of this invention.

[0123] (2) Analysis process The system first performs data quality assessment and modal validity determination on the input data, and filters out modal data whose quality does not meet the preset conditions.

[0124] For valid modes that pass the quality assessment, the system performs feature modeling for each mode and generates a unified motion feature representation in the multimodal feature fusion module.

[0125] Based on this, the system calculates multiple quantitative indicators of movement behavior, including step frequency, symmetry, and stability, and inputs these indicators into the rule-guided analysis module. Under the constraints of preset behavior analysis rules and statistical rules, the system comprehensively judges and corrects the degree of deviation of movement state and risk level.

[0126] (3) Example of analysis results In one exemplary sample, the structured analysis results output by the system include: 1) Multiple quantitative indicators of sports behavior; 2) Overall exercise quality score; 3) A behavioral risk level generated based on the aforementioned score; 4) The corresponding result confidence parameter.

[0127] The analysis results are output in structured data format and can be used for database storage, statistical analysis, batch screening, or analysis report generation. The analysis results of this invention can be uniformly displayed through a visual interface, facilitating centralized viewing and management of motion behavior analysis results from multiple samples.

[0128] To illustrate the organization of analysis results under multi-sample conditions, the system performs batch analysis on multimodal gait data from multiple subjects and presents the results in a visual format.

[0129] like Figure 22 As shown, the system can output structured analysis results for multiple samples, including step frequency, symmetry score, stability score, risk level, and corresponding confidence level. It can also be combined with a rule-guided module to perform logical analysis on the results, providing corresponding analysis prompts and reference suggestions. This demonstrates the batch analysis output of the present invention on several test samples under multimodal input conditions, including structured quantitative indicators (step frequency, symmetry, stability, etc.), fusion weight distribution, and confidence level summary for each sample, facilitating statistical comparison and distribution analysis of the sample set.

[0130] To further illustrate the form of the reasoning results under a single sample condition, in an exemplary embodiment, a representative sample can be selected, its multimodal gait data can be analyzed, and the corresponding results can be displayed.

[0131] like Figure 23 As shown, the system can output structured information, including gait quantification indicators, risk levels, anomaly type identifiers, and corresponding confidence levels, based on multimodal data fusion and indicator-level evaluation results.

[0132] The analysis and reference suggestions are generated by the analysis and processing module. The intelligent analysis module takes the structured analysis results as input, and within the constraints of the statistically constructed behavioral analysis rules and semantic templates, it performs a comprehensive analysis of the indicator combination, abnormal patterns and risk levels, and outputs explanatory text content corresponding to the analysis results.

[0133] The explanatory text is used to explain and prompt the analysis results. Its generation process is constrained by behavioral analysis rules, indicator thresholds and semantic templates to avoid generating unexplainable or irrelevant content.

[0134] (4) Effect description Compared with existing technologies that only output prediction results from a single model, this invention can output more complete and interpretable motion behavior analysis information under multimodal conditions. It not only realizes the hierarchical description of motion states, but also provides index-level evidence for behavioral deviation patterns, demonstrating the technical advantages of the structured motion behavior analysis framework in terms of information integrity and system scalability.

[0135] (II) Example 2: Structured Gait Analysis Process under Single-Mode Conditions (1) Implementation conditions In another implementation, only inertial measurement data is collected as input mode, while other mode data are missing or temporarily unavailable.

[0136] (2) Analysis process The system automatically generates corresponding modality configuration parameters based on the currently available modalities, and performs feature modeling and index calculation processes only for valid modalities.

[0137] For rule judgment items that depend on missing modalities, the system automatically skips or reduces their weight to ensure the continuity and stability of the analysis process.

[0138] (3) Example of analysis results Under these conditions, the system can still output structured analysis results, including gait frequency, stability score, and comprehensive gait quality score, and give the corresponding risk level based on preset risk classification rules.

[0139] When the quality of input data is below the statistical threshold, the system can automatically determine that the sample is invalid input and will not output analysis results.

[0140] (4) Effect description This embodiment demonstrates that the present invention can maintain a complete analysis process and structured output format even under conditions of single modality or lack of modality, thus avoiding the problems of missing analysis results or single dimension caused by insufficient modality in the prior art.

[0141] (III) Example 3: Training Example of a Rule-Based Structured Label Generation Model (1) Implementation conditions In one embodiment, the rule-guided analysis module described in this invention is used to automatically analyze multimodal samples and generate structured labels that include gait quantification indicators, anomaly markers, and analysis confidence levels.

[0142] (2) Training methods The system filters samples based on the confidence parameters of the structured labels, selecting only those samples that meet the preset confidence conditions for model training, thereby reducing the need for manual labeling.

[0143] The model is used to learn gait index prediction and identify movement deviation states, but does not directly output the final diagnostic conclusion.

[0144] (3) Effect description This method allows for the construction of high-quality sample sets for training without the need for manual annotation, achieving model performance close to that of manually annotated training, while also improving the model's adaptability to different modal combinations and data conditions.

[0145] Alternative solutions: Without departing from the overall technical concept of this invention, those skilled in the art can replace, adjust, or extend some implementation methods of the system to adapt to different application requirements or engineering implementation conditions. Such alternative solutions do not affect the overall technical approach and functional logic of this invention in terms of multimodal data organization, fusion analysis, rule constraints, and structured information generation.

[0146] (1) Alternative implementations of multimodal feature fusion methods In an alternative implementation, the multimodal feature fusion module can be replaced or adjusted according to application requirements or computing resource conditions, for example, by adopting weighted fusion, attention mechanism fusion, gating fusion, graph structure fusion or other cross-modal feature integration methods.

[0147] Different fusion methods can correspond to different feature alignment strategies or modal adaptation processes. Their specific implementation forms do not constitute a limitation of this invention. As long as they can effectively integrate multimodal feature information and support subsequent index-level analysis and structured information generation, they fall within the protection scope of this invention.

[0148] (2) Scenario-based configuration of rule-guided and result correction strategies In another alternative implementation, the rule thresholds, weight allocation methods, or logical relationships in the rule guidance module can be configured and adjusted according to specific application scenarios, such as for sports rehabilitation assessment, sports safety risk warning, physical training status analysis, or sports behavior monitoring.

[0149] The above rule configuration adjustments do not change the overall technical approach of this invention, which generates stable and consistent structured analysis results through the synergistic effect of "model analysis results and rule constraints". The specific parameter settings can be flexibly set according to the characteristics of the target population, the collection conditions or the application requirements.

[0150] (3) Expansion and replacement of data modal types and data sources In further alternative embodiments, the system of the present invention can be expanded or replaced with different data modalities, for example, by introducing other types of wearable sensor data, environmental perception data, visual analysis results or physiological signal data, or by using data sources from different devices, different acquisition platforms or different institutions.

[0151] Before being integrated into the system, newly added or replaced data can be adapted based on its physical attributes, time characteristics, or data format to ensure compatibility with existing multimodal analysis processes without requiring substantial modifications to the overall system structure or core functional modules.

[0152] (4) Explanatory Supplement The above alternative solutions are used to illustrate the flexibility and scalability of the present invention under different implementation conditions. Their specific implementation details can be adjusted according to actual engineering needs and do not constitute an essential part of the technical solution of the present invention.

[0153] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. An individual motion behavior analysis system based on dynamic modal configuration and rule guidance fusion, characterized in that, The individual movement behavior analysis and structured information generation system includes: The multimodal data acquisition module is used to acquire multi-source gait-related data of an individual during gait movement or rehabilitation training. The dynamic modality organization and configuration module is used to generate corresponding modality configuration parameters based on the currently available modality types, thereby guiding the subsequent feature modeling, model reasoning and rule analysis processes. The modality configuration parameters are used to indicate the participation status of each modality data in the feature fusion and analysis process, so as to maintain consistent analysis logic and output structure under single-modality or multi-modality input conditions. The multimodal feature modeling and fusion module is used to perform feature modeling on each modality of data and generate a unified gait feature representation. The indicator-level evaluation and behavior state prediction module is used to output multiple quantitative evaluation indicators and movement deviation state identifiers related to gait based on the fused gait feature representation, providing accurate and interpretable quantitative basis for subsequent rule analysis; the multiple quantitative evaluation indicators and movement deviation state identifiers related to gait include at least gait spatiotemporal parameter indicators, symmetry and stability indicators, and multi-label prediction results related to movement deviation state. The rule-guided analysis and result correction module is used to perform rule constraints, logical verification and consistency analysis on indicator-level evaluation results based on the recognized indicator definitions, statistical reference ranges and behavioral evaluation logic in the field of gait biomechanics and motor behavior analysis, and generate motor behavior state analysis results or structured evaluation information to describe behavioral characteristics and their degree of deviation. The structured results generation module is used to uniformly organize indicator-level evaluation results, rule analysis results, and statistical benchmark information to generate a structured motion behavior analysis report that meets the engineering system's requirements for structured data output and multi-scenario reuse.

2. The individual motion behavior analysis system based on dynamic modal configuration and rule-guided fusion according to claim 1, characterized in that, The multimodal data acquisition module is also used to calculate the target data quality using the following formula: in, The proportion of NaN values. The proportion of Inf values, The effective variance is the proportion of channels, and the standard deviation is... The proportion of channels; When the quality of the target data is greater than a preset quality threshold, the target data is determined to be multi-source gait-related data.

3. The individual motion behavior analysis system based on dynamic modal configuration and rule-guided fusion according to claim 1, characterized in that, The dynamic modality organization and configuration module is used to generate corresponding modality configuration parameters based on the currently available modality types, including: The evaluation is conducted across three dimensions: data quality, model consistency, and rule matching. The resulting fusion weights reflect the credibility of the current modality. When the number of available modal types changes, the initial weights of the effective modes need to be dynamically normalized to ensure that the sum of the weights is 1. The formula is as follows: in, These are the normalized modal weights. For the initial weights of the modes, This is the set of currently available modes. To prevent small constants from being divided by zero.

4. The individual motion behavior analysis system based on dynamic modal configuration and rule-guided fusion according to claim 3, characterized in that, The evaluation, which considers three dimensions—data quality, model consistency, and rule matching—outputs fusion weights that reflect the credibility of the current modality. in, For modality Data quality; For modality Model consistency; For modality The rule matching degree is calculated as: the number of rules matched for that modality. The total number of rules associated with this modality; weight coefficients ; ; .

5. The individual motion behavior analysis system based on dynamic modal configuration and rule-guided fusion according to claim 4, characterized in that, The multimodal feature modeling and fusion module is used to perform feature modeling on each modality of data separately and generate a unified gait feature representation, including: A dynamic fusion strategy is adopted, first setting default weights, and then adaptively adjusting them based on data quality and model consistency: The default fusion weights include: Adaptive weight adjustment includes: in, For data quality factors, This is the model consistency factor; The dynamic fusion strategies include: Weights are dynamically switched based on model consensus (Agreement), and consensus is calculated using the following formula: in, For consistency, The dynamic fusion weight is determined based on the pre-defined correspondence between consistency and fusion weight.

6. The individual motion behavior analysis system based on dynamic modal configuration and rule-guided fusion according to claim 5, characterized in that, The indicator-level evaluation and behavior state prediction module is used to output multiple quantitative evaluation indicators and movement deviation state identifiers related to gait based on the fused gait feature representation, providing accurate and interpretable quantitative basis for subsequent rule analysis. The quantitative basis includes: Step 1: For the symmetry index, obtain the comprehensive symmetry score using the following method: 1) Calculate the asymmetry: For different modalities, a unified formula is used to calculate the degree of asymmetry between the left and right sides, as follows: in, To prevent small constants from being divided by zero, V is the core eigenvalue of each mode; 2) Overall symmetry score: To integrate the symmetry information of multiple modalities, a weighted average method is used to fuse the asymmetry of each modality, and then the result is normalized and mapped to a score in the [0,1] interval, as shown in the following formula: in, That is, the weighted average of the asymmetry of each mode; Step 2: For stability and variability indicators, obtain the comprehensive stability score using the following methods: 1) Absolute deviation of the median: It is used to describe the dispersion of data and has strong robustness to outliers. The formula is as follows: in, A time-series data sequence with a certain step-state characteristic. 2) Robust coefficient of variation: The coefficient of variation is calculated based on MAD (Magnitude of Variation) to replace the traditional standard deviation, further improving the tolerance for outliers. The formula is as follows: 3) Window variability: To address the dynamic characteristics of time-series data, the data is divided into multiple windows, and the degree of feature variation between windows is calculated to reflect the stability of gait rhythm. Window mean calculation: The default value is N=4, and the window size is 128 / 4=32 frames. Quantification of variability: ; 4) Overall stability score: After weighted fusion of multimodal variability, it is transformed into a score in the [0,1] interval through nonlinear mapping, as shown in the following formula: in, The weighted average of modal variability. Step 3, obtain the step frequency in the following way: 1) Signal preprocessing The IMU data is mean-centered to remove baseline drift, using the following formula: in, For IMU time series data, The global mean of the IMU data; 2) Normalized autocorrelation The autocorrelation function of the signal is calculated to capture the periodicity of the gait cycle, as shown in the following formula: in, For the lagging step size, Frame (corresponding to approximately 180 steps / minute) Frames (corresponding to approximately 40 steps / minute) cover the step frequency of gaits falling within the statistical reference range; 3) Period detection and step frequency calculation Find the first peak of the autocorrelation function within the effective lag range [10, 60], which corresponds to the gait period. The formula is as follows: The step frequency is calculated based on the period, where FPS=30, using the following formula: The effective range of cadence is [40, 180] steps / minute. If it exceeds this range, it is judged as an abnormal gait rhythm.

7. The individual motion behavior analysis system based on dynamic modal configuration and rule-guided fusion according to claim 6, characterized in that, The rule-guided analysis and result correction module is used to perform rule constraints, logical verification, and consistency analysis on indicator-level evaluation results based on the generally accepted indicator definitions, statistical reference ranges, and behavioral evaluation logic in the field of gait biomechanics and motor behavior analysis. It also generates motor behavior state analysis results or structured evaluation information describing behavioral characteristics and their degree of deviation, including: Taking into account three factors—data quality, model consistency, and rule consistency—the overall confidence level of the analysis results is calculated using the following formula: in, Indicates the degree of consistency between different analysis paths; Indicates consistency of rules; Confidence of pseudo-labels: in, Indicates sample consistency. Indicates the rule matching degree.

Images