Gait monitoring system based on multi-modal sensing fusion

By using multimodal sensor fusion and machine learning to dynamically adjust thresholds, the system addresses the issues of insufficient adaptation to individual differences, data reliability, and multi-timescale analysis in existing gait monitoring systems, thus achieving efficient gait monitoring and rehabilitation assessment.

CN121313148APending Publication Date: 2026-01-13WUXI NO 2 PEOPLES HOSPITAL +1

Patent Information

Application Number
CN202511491270.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-10-18
Publication Date
2026-01-13

AI Technical Summary

Technical Problem

Existing gait monitoring systems lack the ability to adapt to individual differences, cannot dynamically adjust monitoring thresholds, have insufficient data reliability, lack multi-timescale analysis, and lack disease specificity, resulting in a high misjudgment rate and an inability to accurately reflect rehabilitation progress.

Method used

By employing multimodal sensor fusion technology, combining accelerometers, gyroscopes, and pressure sensors, an individual gait baseline is established through machine learning, thresholds are dynamically adjusted, and multi-layered abnormal data filtering and multi-level adaptation mechanisms are used to achieve multi-timescale analysis and disease-specific monitoring.

Benefits of technology

It achieves individualized adaptation, improves the accuracy and robustness of monitoring, enables rapid short-term response and long-term tracking of rehabilitation trends, and provides personalized rehabilitation guidance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121313148A_ABST
    Figure CN121313148A_ABST
Patent Text Reader

Abstract

The invention provides a gait monitoring system based on multi-modal sensing fusion, which comprises the following steps of: establishing an initial base line to form a personalized gait base line model; baseline updating is adopted, and after a system obtains new gait data, a weighted moving average algorithm is adopted for self-learning updating, so that the historical data weight is enhanced, and the baseline stability is improved. The technical problems of insufficient individual difference adaptation, lack of dynamic threshold adjustment, poor real-time performance and the like in the prior art are solved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of rehabilitation equipment technology, specifically relating to a gait monitoring system based on multimodal sensor fusion. Background Technology

[0002] With the increasing aging population and rising incidence of neurological diseases, the demand for gait rehabilitation training is growing. As described in patent 202511078923.3, a gait monitoring and muscle strength monitoring fusion algorithm and a rehabilitation walking aid using the same, a gait monitoring algorithm includes: a gait feature extraction module for extracting gait feature values ​​from the raw sensor data, including: cadence, calculated by analyzing periodic fluctuations in pressure or acceleration to determine the number of steps per second; and stride length, calculated by integrating acceleration or the interval between the left and right feet touching the ground to determine the distance between two steps. However, this algorithm has the following technical shortcomings: Fixed threshold issue: Existing systems often use fixed gait parameter thresholds for anomaly detection, failing to consider individual differences. Different users exhibit significantly different normal gait characteristics due to factors such as age, height, weight, and disease type. Fixed thresholds lead to oversensitivity for some users (high false positive rate) and insufficient sensitivity for others (high false negative rate).

[0003] Lack of adaptive capability: The gait characteristics of rehabilitation patients continue to improve as the rehabilitation process progresses, and the existing system cannot dynamically adjust the monitoring criteria, thus failing to accurately reflect the user's rehabilitation progress.

[0004] Data reliability issues: Sensor data is susceptible to environmental interference, usage habits, sensor malfunctions, and other factors. Existing systems lack effective abnormal data filtering mechanisms, leading to misjudgments based on low-quality data.

[0005] Single time scale: Existing systems typically perform analysis only on a single time scale, which cannot simultaneously take into account both short-term fluctuation response and long-term trend tracking.

[0006] Lack of disease specificity: Different diseases (such as stroke, Parkinson's disease, etc.) have different gait characteristics and rehabilitation patterns. Existing systems mostly use general algorithms and have failed to be optimized for specific diseases.

[0007] Therefore, there is an urgent need to develop an intelligent gait monitoring system that can automatically learn individual gait characteristics, dynamically adjust monitoring thresholds, effectively filter abnormal data, and take into account multi-timescale analysis. Summary of the Invention

[0008] The purpose of this invention is to provide a gait monitoring system based on multimodal sensor fusion, which solves the technical problems of insufficient adaptation to individual differences, lack of dynamic threshold adjustment, and poor real-time performance in the prior art.

[0009] The existing hardware and software, which follows the format of 202511078923.3, will not be elaborated upon here.

[0010] Accelerometer (IMU) / Gyroscope (Gyro) × 2; Installation location: Middle of the two front uprights (near knee height). Function: Used to monitor the patient's overall direction of movement, acceleration and speed changes, reflecting walking rhythm and posture deviations.

[0011] Data type: Triaxial acceleration (Ax, Ay, Az) + Triaxial angular velocity (Gx, Gy, Gz); 2. Accelerometer (IMU) / Gyroscope (Gyro) × 2; Installation location: Integrated with the acceleration sensor in the same module (combined IMU), also located in the middle of the two rear columns; Function: Real-time acquisition of the patient's trunk and upper body rotation angle and tilt, to determine whether the body is leaning forward or to the side.

[0012] Data type: Triaxial acceleration (Ax, Ay, Az) + Triaxial angular velocity (Gx, Gy, Gz), 1 and 2 are combined into a 6-axis sensor.

[0013] 3. Pressure sensors × 6; Installation location: × 2 below the handle: Detects the symmetrical application of force by both hands; Four pillars at their base × 4: Determine the contact pressure of the left, right, front, and back pillars on the ground to analyze the patient's center of gravity shift and standing stability; Function: To determine whether both hands exert force evenly, and to help analyze whether there is hemiplegia; Detect whether gait rhythm and cadence are consistent; Determine the time distribution of the support period and the difference in weight-bearing between the left and right feet.

[0014] One of the purposes of data training is: For cadence training, the cadence should be maintained within a reasonable range (e.g., 0.8~1.5Hz). Stride length training should aim to achieve a minimum stride length (e.g., ≥ 40cm). For body posture training, the upper body should be kept upright with an offset angle of ≤ 10°.

[0015] These raw data are preprocessed by algorithms and transformed into the gait features we need.

[0016] II. Core "feature values" extracted from raw sensor data.

[0017] Feature name, meaning, and calculation method.

[0018] Step frequency (Hz) is the number of steps taken per second. It is estimated by the periodic fluctuations in pressure or acceleration, with a typical range of 0.8 to 1.5 Hz.

[0019] Stride length (m): the distance between two steps. Displacement is obtained by integrating acceleration or calculated from the distance between the left and right feet touching the ground. The target value is generally >0.4m.

[0020] Body forward tilt angle, pitch angle. Calculated by integrating the angular velocity from the gyroscope or by directly reading the attitude (usually using Gx data).

[0021] Right pressure difference (N), the pressure difference on the handle or column; Gait rhythm changes, fluctuations in continuous gait frequency or stride length (such as standard deviation), are used to determine stability. This is employed for sliding window anomaly detection.

[0022] Ground contact duration, the duration of continuous contact between each foot and the ground, is derived from the rhythm of the pressure sensor. Abnormalities are indicated by one side having an excessively short contact time. (Please note that this sampling value can only be collected by standard four-legged walking aids; it cannot be collected by walking aids with front wheels for the following reasons: the front wheels do not sense vertical pressure during walking, and the rear feet are generally covered by anti-slip pads, potentially only contacting the ground for a limited period. Therefore, the left and right foot support information may be masked in the rhythm of such walking aids. Forcibly sampling this type of walking aid will not reveal the user's walking pattern.) Personal baseline establishment: When a user first uses the system, it establishes a "personal normal gait pattern" based on several normal walking records. For example, if a user's normal cadence is found to be 0.95Hz, then their personalized threshold should be 0.85-1.05Hz, not the generic 0.8-1.5Hz. A well-designed gait system should have a false alarm rate below 5%. However, without individual users, the false alarm rate could be as high as 20-30%.

[0023] Historical data learning: The system should remember a user's gait performance over the past week or month, noting that "Mr. Zhang's cadence has recently improved from 0.8Hz to 1.1Hz," and then adjust expectations accordingly. Small improvements are crucial during rehabilitation. For example, an improvement in stride length from 0.35 meters to 0.38 meters, while still below the "standard" of 0.4 meters, is significant for this patient. Systems lacking a personalized core cannot recognize and encourage such progress, failing to reflect true improvement.

[0024] Poor user experience: Environmental adaptation: The system should be able to recognize that "it is normal for the user to walk outdoors today and have different gait parameters than indoors", rather than simply issuing an alarm.

[0025] Therefore, the algorithm is improved as follows: User Learning Module: Add a dedicated algorithm module that describes how to establish a personal gait baseline using machine learning methods, including data collection strategies, baseline update algorithms, outlier filtering mechanisms, etc.

[0026] Dynamic threshold calculation method: describes how to calculate thresholds in detail based on personal historical data, such as a specific formula like "personal step frequency threshold = recent ±1.5 times standard deviation".

[0027] Multi-level adaptation mechanism: Design a tiered adaptation strategy that includes short-term adaptation (intraday adjustment), long-term adaptation (weekly adjustment), and long-term adaptation (adjustment of the recovery process).

[0028] Data collection strategy designed in stages: The initial phase employs a gradual data collection strategy. During the first three uses, the system collects 5-10 minutes of continuous gait data each time. This duration is sufficient to capture the user's natural gait patterns without causing fatigue. In this phase, the system guides the user through standardized gait tests, including normal-speed walking, slightly faster walking, and turning movements, to gain a complete description of the user's gait capabilities.

[0029] The data quality control mechanism ensures that representative information is collected. The system automatically identifies and corrects abnormal data, such as user initial nervousness, accidental trips, or pauses. By analyzing the continuity and consistency of signals, the system can filter out data segments that truly represent the user's normal gait.

[0030] The core update algorithm for mathematical modeling: The individual gait core employs a weighted moving average algorithm. The specific mathematical expression is: New baseline parameter = α × current measurement + (1-α) × historical baseline parameter The learning rate α is dynamically adjusted based on the reliability of the data. The α value for new users is set to 0.3, indicating a faster learning speed; after one month of stable use, the α value drops to 0.1, indicating a greater emphasis on the stability of historical data.

[0031] The comprehensive modeling of multidimensional features takes into account the favorable variations of multiple factors such as gait frequency, stride length, body posture, and pressure distribution. The system uses principal component analysis to identify the intrinsic relationships between these parameters and establish a multidimensional mathematical model of individual gait. In this way, when some parameters change, the system can predict the reasonable range of changes in other parameters.

[0032] Intelligent mechanisms for outlier filtering: The third-layer anomaly detection mechanism ensures the reliability of baseline updates. The first layer is a rapid screening based on statistical thresholds; if a measured cadence deviates from an individual's historical mean by more than three standard deviations, the system will mark it as suspicious data. The second layer is verification based on physiological plausibility; for example, a user's cadence is unlikely to suddenly increase from 1.0Hz to 1.8Hz in a single day. The third layer is analysis based on temporal continuity; if abnormal data persists for more than three days, the system will consider this a change in the user's actual condition rather than a data error.

[0033] Machine learning-driven anomaly detection employs the Isolation Forest algorithm, which can identify complex, multidimensional anomaly patterns. This approach is particularly well-suited for processing data from rehabilitation patients, whose gait often exhibits complex, non-linear patterns.

[0034] Accurate implementation of the dynamic threshold calculation method: Personalized threshold calculations require finding the optimal balance between individual residential needs and clinical safety. We cannot allow thresholds to be too high, thus missing genuine abnormalities, nor can we be too strict, resulting in a large number of error reports.

[0035] Formula for calculating personal cadence threshold: Upper limit of personal cadence threshold = Personal baseline cadence + 1.5 × Personal cadence standard deviation + Safety margin Lower limit of personal cadence threshold = Personal baseline cadence - 1.2 × Personal cadence standard deviation - Safety margin The coefficients 1.5 and 1.2 here represent the most frequent values ​​validated by extensive clinical data. The upper limit is increased because walking too fast is generally less risky than walking too slow, while the lower limit is decreased to detect gait deviations early. The safety margin is dynamically adjusted based on the user's disease type and rehabilitation stage; a larger safety margin is set for Parkinson's disease patients to accommodate the disease's fluctuating characteristics.

[0036] Consistent threshold optimization across multiple parameters: The calculation of stride threshold needs to consider its coordination with stride frequency: Personal stride threshold = f(height, leg length, baseline cadence, historical stride pattern); Dynamic adjustment factor = Current step frequency / Baseline step frequency; Adjusted stride threshold = Base stride threshold × Dynamic adjustment coefficient^0.7; This index of 0.7 comes from biomechanical research on the relationship between cadence and stride length, reflecting the natural stride length adjustment pattern of the human body at different cadences.

[0037] Introduction of time decay weight: To ensure that the threshold reflects both the user's long-term state and adapts to short-term changes, we introduce a time decay weight: Weight(t) = exp(-λ × (current time - data time)); Weighted average = Σ(data value × weight) / Σ(weight); Here, λ is the attenuation coefficient, set to 0.1 / day, meaning that the weight of data from a week ago is reduced to about 50% of the current data. This maintains the reference value of historical data while ensuring that recent changes are accounted for.

[0038] Hierarchical design of multi-level adaptation mechanisms: Gait monitoring needs to be adjusted at different time scales. Just like rapid adjustments, the body's physiological regulatory mechanisms also have slow changes in sleep.

[0039] Short-term adaptation mechanism (intraday adjustment): During a single use, the system makes real-time micro-adjustments. For example, a user might initially exhibit a restrained gait, gradually relaxing to a natural state after six months. The system monitors this change through a sliding window. Real-time adjustment threshold = Base threshold × (1 + Intraday variation coefficient) Intraday variation coefficient = tanh((current time - start time) / adaptation time constant) This hyperbolic tangent function ensures smooth adjustment, avoiding sudden jumps in the threshold. The adaptation time is set to 15 minutes by default, reflecting the time most users need to transition from a stressed state to a natural state.

[0040] Recognizing user fatigue is also an important part of short-term adaptation. When the system detects a gradual decrease in cadence and an increase in gait variability, it will relax the threshold criteria accordingly to avoid issuing inappropriate "increase cadence" prompts when the user is fatigued.

[0041] Periodic adaptation mechanism (weekly adjustment): The system performs a comprehensive evaluation and adjustment weekly. This process analyzes the user's gait data patterns over the course of a week to identify systemic trends. Weekly adjustment coefficient = Linear regression slope × Confidence weight + Volatility correction; New weekly baseline = Last week's baseline × (1 + Weekly adjustment factor); The linear regression slope reflects the improvement or decline in a user's gait ability, and the confidence weights ensure that only statistically significant changes trigger threshold adjustments. Volatility modifications take into account data stability; if a user's performance fluctuates significantly within a week, the adjustment magnitude will be reduced accordingly.

[0042] Users tend to have a stronger gait in winter, and the system will learn this pattern and adjust the expected value accordingly.

[0043] Long-term adaptation mechanisms (adjustment of the recovery process): Understanding the adaptive mechanisms of the rehabilitation process requires comprehending the physiological patterns of disease recovery. Different diseases have specific recovery time curves: Stroke rehabilitation follows an exponential decay model: Expected improvement (t) = Maximum improvement potential × (1 - exp(-t / recovery time constant)); Recovery time varies from person to person, typically ranging from 3 to 6 months. The system will dynamically adjust this parameter based on the user's specific situation.

[0044] The adaptive mechanisms of Parkinson's disease need to take into account the progressive characteristics of the disease: Disease progression correction = Baseline rate of decline × Time × Individual progression coefficient; Adjusted threshold = Standard threshold - Disease progression correction + Training improvement; The design allows the system to find this balance between the natural progression of the disease and the effectiveness of rehabilitation training, providing users with realistic yet encouraging goal setting.

[0045] Multi-system coordination mechanism: The mechanisms of the three systems need to work in coordination and are in conflict with each other. The system adopts a priority strategy: safety considerations have the highest priority, followed by long-term rehabilitation goals, and short-term comfort adjustments have the lowest priority.

[0046] When adjustments to different systems conflict, the system employs a weighted fusion approach: Final threshold = w1 × short-term threshold + w2 × medium-term threshold + w3 × long-term threshold; Constraints: w1 + w2 + w3 = 1, and the lower safety bound is satisfied; The weighting is dynamically adjusted based on the user's recovery stage. The "Rehabilitation-Based" approach prioritizes long-term goals, while the "Rehabilitation-Based" approach focuses more on short-term comfort.

[0047] This layered design ensures that the system can respond to immediate changes in the user's status, quickly serve advanced long-term rehabilitation goals, and maintain sufficient security and stability. Through the complete implementation of these three modules, this application will possess truly personalized monitoring capabilities, achieving significant improvements in both technical aspects and practical value. Beneficial effects

[0048] Compared with the prior art, the present invention has the following beneficial effects: Individualized adaptation: Through the establishment of an initial baseline and a continuous self-learning mechanism, the system can automatically adapt to the unique gait characteristics of each user, avoiding misjudgment problems caused by fixed thresholds.

[0049] High robustness: A four-layer anomaly filtering mechanism (statistical screening, physiological rationality verification, temporal continuity verification, and isolated forest algorithm) ensures data credibility and significantly improves monitoring accuracy.

[0050] Multi-timescale adaptation: The three-level adaptive mechanism of short-term, periodic and long-term works together to quickly respond to short-term fluctuations and accurately track long-term recovery trends.

[0051] Parameter co-optimization: Considering the biomechanical coupling relationship between stride frequency and stride length, multi-parameter co-correction is achieved to improve the rationality of threshold calculation.

[0052] Disease specificity: Specific rehabilitation curve models are used for different diseases such as stroke and Parkinson's disease to provide accurate rehabilitation assessments.

[0053] Intelligent conflict coordination: When multiple mechanisms conflict, they are automatically integrated and adjusted to ensure user safety with a safety-first principle.

[0054] Real-time feedback: Personalized voice prompts provide instant guidance, improving adherence to and effectiveness of rehabilitation training. Attached Figure Description

[0055] Figure 1 System overall architecture diagram Figure 2 Flowchart of Multi-Level Outlier Filtering Mechanism Figure 3 Dynamic threshold calculation flowchart Figure 4 Schematic diagram of multi-level adaptive mechanism Figure 5 Flowchart of Multi-Mechanism Conflict Coordination Algorithm Detailed Implementation

[0056] refer to Figure 1-5 Example 1: Individualized learning and baseline establishment.

[0057] Step 1: Data Collection; As above.

[0058] Step 2: Initial Baseline Establishment When users first use the system, they are required to complete three independent walking training sessions in a normal walking environment, each lasting 5-10 minutes. The system automatically extracts the following gait features: (1) Step frequency calculation: The gait period is identified by detecting the periodic peak value of the acceleration signal. The step frequency f is defined as the number of steps per unit time: f = N_number of steps / T_total number; Where N_steps is the number of steps, T_total is the total time (in seconds), and the step frequency is in Hz.

[0059] (2) Stride length estimation: Based on the acceleration integral and step frequency information, the stride length L is estimated: L = v_avg / f; Where v_avg is the average walking speed, which is estimated by double integral of acceleration combined with the Zero Velocity Update (ZUPT) algorithm.

[0060] (3) Calculation of body forward tilt angle: Based on accelerometer and gyroscope data, the forward tilt angle θ of the body relative to the vertical direction is calculated by using complementary filter or Kalman filter fusion. θ = arctan(ax / sqrt(ay²+az²)); (4) Pressure symmetry index: Defined as the ratio of the pressure on the left side to the pressure on the right side: SI = (PL + P1) / (PR + P2); Ideally, SI=1 indicates perfect symmetry.

[0061] The above features were calculated for each of the three training datasets, and then the average value was taken as the baseline parameter. Baseline step rate f_baseline = mean(f1, f2, f3); Baseline stride L_baseline = mean(L1, L2, L3); Baseline forward tilt angle θ_baseline = mean(θ1, θ2, θ3); Baseline symmetry SI_baseline = mean(SI1, SI2, SI3); Simultaneously, the standard deviation σ of each parameter is calculated for subsequent threshold calculation.

[0062] Step 3: Baseline dynamic update; In subsequent use, after each training session, the system updates the baseline using a weighted moving average algorithm: f_baseline_new = α × f_current + (1-α) × f_baseline_old; in: Initial phase (use days ≤ 30): α = 0.3, rapid adaptation to individual characteristics; Stable phase (usage days > 30 days): α = 0.1, increase the weight of historical data; This mechanism ensures that the baseline responds to real gait improvements without changing frequently due to short-term fluctuations.

[0063] Results: This method can continuously learn the individual gait change trend of users during rehabilitation training, realize adaptive modeling of personal gait characteristics, and reduce the false alarm and false negative rates of fixed threshold algorithms.

[0064] Example 2: Multi-layer outlier filtering; First layer: Statistical filtering; For each gait data collection, calculate the degree of deviation from the historical mean: Deviation = |x_current - μ_history| / σ_history; Data with a deviation greater than 3 is marked as suspicious.

[0065] For example: A user's historical average step frequency is 1.2Hz, with a standard deviation of 0.08Hz. If a certain measurement shows a step frequency of 1.5Hz, then: Deviation = |1.5 - 1.2| / 0.08 = 3.75 > 3; The data was flagged as suspicious.

[0066] The second layer: verification of physiological rationality; Establish physiological boundaries for human gait parameters: Step frequency range: 0.5 - 2.5 Hz; Stride range: 20 - 100 cm; Lean angle range: 0 - 35°; Symmetry index range: 0.5 - 2.0; Limit on parameter change rate within 24 hours: Step frequency change rate <30%; Stride change rate <40%; Forward tilt angle change <10°; For example: A user's cadence was 1.0Hz yesterday and was measured at 1.8Hz today, a change rate of 80%, which is outside the reasonable range, so this data is discarded.

[0067] Third layer: Time continuity verification; Use a sliding time window (7 days) to count the frequency of outliers: Anomaly frequency = N_anomaly / N_total; A change is considered a functional change rather than abnormal noise when the following conditions are met: Abnormal frequency >50%; The duration is ≥ 3 days; The trend of change is monotonous (continuously rising or falling); This triggers the baseline update process.

[0068] Fourth layer: Isolation Forest algorithm; Combining gait features into a multi-dimensional vector: X = [f, L, θ, SI, CV]; Where CV is the gait variation coefficient.

[0069] Anomaly detection using the Isolation Forest algorithm: Construct 100 isolation trees, and randomly select features and split points for each tree; Calculate the average path length h(x) of the sample in each tree; Calculate the outlier score: s(x) = 2^(-h(x) / c); Set a threshold th=0.6, and determine an anomaly when s(x)>th.

[0070] Forest isolation can identify complex anomalous patterns in multidimensional space, such as "normal cadence but abnormally small stride and excessive forward lean angle".

[0071] Results: Through the above three-layer filtering and algorithm detection mechanism, the system achieves highly robust data screening in gait monitoring, ensuring the reliability of baseline updates and the accuracy of rehabilitation assessment results.

[0072] Example 3: Dynamic Threshold Calculation Step frequency threshold calculation Based on individualized baselines and statistical dispersion, an asymmetric thresholding strategy is employed: Upper limit of step frequency = f_baseline + 1.5 × σ_f + M_upper Step frequency lower limit = f_baseline - 1.2 × σ_f - M_lower Where M is the safety margin, adjusted according to the type of disease: Disease type: Early stroke; M_up: 0.05; M_lower: 0.15; Note: Strictly monitor for decreased cadence; Disease type: Late stroke; M_up: 0.10; M_lower: 0.10; Description: Balance monitoring; Disease type: Parkinson's disease; M_upper: 0.05; M_lower: 0.10; Description: To prevent panic gait; Disease type: Elderly health; M_upper: 0.08; M_lower: 0.08; Explanation: Maintain within the normal range.

[0073] For example: For a stroke patient in the early stages, f_baseline = 1.0Hz, σ_f = 0.1Hz, then: Upper limit of step frequency = 1.0 + 1.5 × 0.1 + 0.05 = 1.20 Hz; Lower limit of step frequency = 1.0 - 1.2 × 0.1 - 0.15 = 0.73 Hz; Stride length coordination correction: Stride length and gait frequency are coupled in human gait, following a power law: L ∝ f^β; Where β≈0.7. Based on this, when the real-time cadence deviates from the baseline, the stride threshold is automatically corrected: L_threshold_adjusted = L_threshold_base × (f_current / f_baseline)^0.7; For example: with a base stride threshold of 60cm, a baseline cadence of 1.0Hz, and a current cadence of 1.2Hz: The adjusted stride threshold = 60 × (1.2 / 1.0)^0.7 = 60 × 1.14 = 68.4 cm; This allows the threshold to respond to changes in step frequency, avoiding misjudging excessive stride length due to increased step frequency.

[0074] Dynamic parameter binding: When a change in stride length or forward lean angle is detected, the system automatically recalculates the relevant parameter thresholds: IF step size increase >10% THEN; The upper limit for step frequency has been relaxed by 5%. The lower limit of the camber angle has been tightened by 5%. END IF.

[0075] This coupling correction ensures consistency in multi-parameter monitoring.

[0076] Results: This method can adapt to the user's physiological characteristics and training stage, significantly reduce misjudgments caused by static thresholds, and support accurate gait trend tracking across time periods.

[0077] Example 4: Time decay weight; Weight function design; Assign exponentially decaying weights to historical data: w(t) = exp(-λ × Δt); Where Δt is the number of days since the data was collected, and λ = 0.1 / day.

[0078] Weight example:

[0079] Calculation of weighted statistics; Weighted average: f_weighted = Σ(fi × wi) / Σ(wi); Weighted standard deviation: σ_weighted = sqrt(Σ(wi × (fi - f_weighted)²) / Σ(wi)); Adaptive λ adjustment; When an increase in gait variability is detected (increased coefficient of variation CV), the λ value is automatically increased, causing historical data to decay more quickly and improving sensitivity to recent data. λ_adjustment = λ_base × (1 + 0.5 × ΔCV); Where ΔCV is the increment of the coefficient of variation relative to the baseline.

[0080] For example: if the baseline coefficient of variation is 15%, and the current coefficient of variation rises to 25%, then ΔCV = 10%. λ_adjustment = 0.1 × (1 + 0.5 × 0.1) = 0.105 / day.

[0081] Results: This method enables a smooth transition of gait parameters over time, making the trend analysis of individual gait changes more continuous and interpretable, and effectively improving the temporal stability of personalized rehabilitation monitoring.

[0082] Example 5: Multi-level adaptive mechanism; Short-term adaptation (intraday adjustment); In a single use, a hyperbolic tangent function is used to achieve a smooth transition: k(t) = tanh((t - t0) / τ); in: t0 = 30 seconds (adaptation start point); τ = 120 seconds (time constant); Real-time threshold correction: Threshold_Real-time = Threshold_Baseline × (1 + 0.15 × (1 - k(t))); In the initial stage of use (the first 2 minutes), k(t) is relatively small, and the threshold is relaxed by about 15%; after 2 minutes, k(t) approaches 1, and the threshold returns to the baseline value.

[0083] Fatigue detection: Fatigue is determined when the following conditions are met simultaneously: The average cadence over the last 5 minutes is less than 0.92 times the baseline. Gait variation coefficient > baseline × 1.3; Symmetrical fluctuation > 0.15; When fatigued, the lower threshold is automatically relaxed by 10%, and a voice prompt is given to take a rest.

[0084] Periodic adaptation (weekly adjustment); A retrospective analysis is conducted every 7 days. Collect data from the past 7 days and remove outliers; Fitting a linear trend: y = kt + b; Calculate the R² coefficient of determination to determine the significance of the trend; If R² > 0.6 and |k| > 0.05 × baseline, then adjust the baseline: New baseline = Current baseline × (1 + k × 7 days / Baseline × Adjustment factor); The adjustment factor is determined based on the volatility: Low volatility (CV < 15%): Adjustment factor = 0.5 (fully adopting the trend); Large volatility (CV>25%): Adjustment factor = 0.2 (conservative adjustment); Long-term adaptation (adjustment of the rehabilitation process); For stroke patients, an exponential recovery model is used: f(t) = A × (1 - exp(-t / τ)); in: A = Maximum potential for improvement (based on initial assessment); τ = recovery time constant, typically 60-90 days; t = Number of days since the start of rehabilitation; Assess the agreement between actual progress and model predictions every 30 days and adjust the baseline accordingly. If actual improvement > expected improvement × 1.2 then; Increase the baseline by 5-10% (for better recovery). ELSE IF Actual improvement < Expected improvement × 0.8 THEN; Maintain the baseline or slightly reduce it (recovery encounters obstacles). END IF.

[0085] For Parkinson's disease patients, a linear progression model was used: Expected gait (t) = Baseline - Disease progression rate × t + Improvement from training; Determining disease progression rate based on HY staging: Early stage (HY I-II): 1-2% / year; Mid-term (HY III): 3-5% / year.

[0086] Results: Through a layered time-adaptive design, the system can respond to short-term fluctuations in real time and track rehabilitation trends in the long term, achieving dynamic optimization of individuals throughout the entire cycle.

[0087] Example 6: Multi-mechanism conflict coordination Example of a conflict scenario; Scenario 1: Fatigue is detected in the short term (it is recommended to relax the threshold), but the weekly trend shows improvement (it is recommended to tighten the threshold). Scenario 2: Weekly data fluctuates greatly (conservative adjustment is recommended), but long-term model predictions should improve (active adjustment is recommended). Weighted fusion algorithm; Final threshold = w1 × short-term threshold + w2 × periodic threshold + w3 × long-term threshold; Weights are allocated based on the stage of rehabilitation: Explanation of rehabilitation phases w1, w2, and w3; Initial phase (<2 weeks): 0.5 0.3 0.2 Emphasis on immediate feedback; Rapid recovery period (2 weeks - 3 months): 0.2 0.5 0.3 (Important: Weekly) Progress and stabilization period (3-6 months): 0.1 0.4 0.5 Emphasis should be placed on long-term patterns; Maintenance period (>6 months): 0.15 0.35 0.5 Balanced stability and sensitivity; Dynamic weight adjustment; Adjust weights when a special state is detected: IF Fatigue detected THEN w1 = w1 × 1.5 (increase short-term weight); Renormalization: w1 + w2 + w3 = 1; END IF; IF stable progress detected THEN; w3 = w3 × 1.3 (increase long-term weight); Renormalization; END IF.

[0088] Safety first rule; When conflicts still exist in the fusion results, a decision tree is used: IF any mechanism recommends tightening the threshold (increasing alert) THEN; Use the most stringent threshold; ELSE IF any mechanism detects a security risk THEN; Use the most lenient threshold (reduce training intensity); ELSE IF if the difference between the mechanisms is >20% THEN; Use the median of the three; ELSE; Use weighted fusion results; END IF.

[0089] Results: The algorithm ensures that multi-time-level models operate collaboratively, achieving a balance between safety, comfort, and rehabilitation goals, and ensuring that the system output thresholds are stable, reasonable, and individualized.

Claims

1. A gait monitoring system based on multimodal sensor fusion, characterized in that, Includes the following steps: Data acquisition phase: Raw gait signal data of the user is acquired by a multimodal sensor installed on the walker. The multimodal sensor includes a 6-axis accelerometer and gyroscope module and four bottom pressure sensors. The data includes three-axis acceleration, three-axis angular velocity and handle pressure on the ground. Initial baseline establishment phase: The system collects continuous walking data for at least three 5-10 minute sessions when the user first uses the system. By extracting cadence, stride length, body lean angle and pressure symmetry features, the system calculates the user's average gait parameters and forms a personalized gait baseline model. Baseline update phase: After the system acquires new gait data, it uses a weighted moving average algorithm for self-learning updates. The mathematical model is as follows: New baseline parameter = α × current measurement + (1-α) × historical baseline parameter Here, α is the learning rate coefficient, initially set to 0.3, and automatically reduced to 0.1 after the user's usage period exceeds 30 days, in order to enhance the weight of historical data and improve baseline stability.

2. The gait monitoring system based on multimodal sensor fusion according to claim 1, characterized in that, The system further includes a multi-layered outlier filtering mechanism to screen the sampled data for reliability before baseline updates, specifically including: The first layer is a statistical screening mechanism. When any gait feature value (step frequency, stride length, or posture angle) deviates from the user's historical mean by more than three times the standard deviation, the system marks the data as suspicious. The second layer is a physiological rationality verification mechanism to determine whether the data changes exceed the reasonable range of the human body. For example, the cadence cannot suddenly increase from 1.0Hz to 1.8Hz within 24 hours; otherwise, it will be eliminated. The third layer is the time continuity verification mechanism. When outliers of the same feature continue to appear for more than three days, the system determines that they are functional changes rather than abnormal noise and automatically triggers baseline updates. Intelligent detection algorithm layer: The system uses the Isolation Forest algorithm to perform unsupervised learning on multi-dimensional gait feature vectors, automatically identify complex nonlinear anomaly patterns, and distinguish between real anomalies and sensor interference data.

3. The gait monitoring system based on multimodal sensor fusion according to claim 1 or 2, characterized in that, The system includes a dynamic threshold calculation module, used to generate personalized cadence and stride length thresholds in real time on an individual's baseline, specifically including: Step frequency threshold adaptive formula: Upper limit threshold = Baseline cadence + 1.5 × Standard deviation of cadence + Safety margin Lower threshold = Baseline cadence - 1.2 × Standard deviation of cadence - Safety margin The safety margin is automatically adjusted based on the type of disease (such as stroke, Parkinson's disease) and the stage of recovery to balance oversensitivity and the risk of missed detection. Stride length cooperative correction model: Adjust stride threshold = Baseline stride threshold × (Current stride frequency / Baseline stride frequency)^0.7 The index 0.7 comes from the biomechanical coupling law between cadence and stride, enabling the threshold to dynamically respond to changes in the user's cadence. Parameter dynamic binding mechanism: When stride or attitude angle changes, the system automatically recalculates the range of interdependent parameters to achieve coupling correction between multiple features.

4. The gait monitoring system based on multimodal sensor fusion according to claim 3, characterized in that, To ensure the system balances sensitivity to recent changes with historical stability, a time decay weighting mechanism is introduced, the formula of which is as follows: Weight(t) = exp(-λ × (current time - data time)) Wherein, λ is the time decay coefficient, preferably taken as 0.1 / day, so that the weight of data from one week ago decays to about 50% of the current weight; The weighted fusion of gait data from multiple time periods is calculated using the following formula: Weighted average = Σ(data value × weight) / Σ(weight) In addition, the λ value is dynamically adjusted in conjunction with the anomaly detection module, and the recent weight is automatically increased when the system detects an increase in volatility.

5. The gait monitoring system based on multimodal sensor fusion according to any one of claims 1 to 4, characterized in that, The system includes a three-level adaptive mechanism—short-term, periodic, and long-term—to dynamically adjust individual gait models at different time scales. Short-term adaptation mechanism: During a single use, the system analyzes step frequency fluctuations in real time through a sliding time window and corrects the threshold using a hyperbolic tangent function. Real-time adjustment coefficient = tanh((current time - starting point) / adaptation constant); It responds smoothly to changes in stress or fatigue during the initial stages of use; Periodic adaptation mechanism: The system performs a comprehensive review every 7 days, and calculates the weekly adjustment coefficient based on the linear regression slope and confidence level of one week's gait data. New baseline = Last week's baseline × (1 + adjustment factor); And reduce the adjustment range when gait fluctuations are large to prevent erroneous updates; Long-term adaptation mechanism: The system adopts a rehabilitation curve model for different rehabilitation populations. Stroke rehabilitation follows an exponential decay model: f(t) = Maximum improvement potential × (1 - e^(-t / recovery time constant)) Parkinson's disease patients were treated using a linear progression modified model: Threshold = Standard Threshold - Progress Correction + Training Improvement 6. The gait monitoring system based on multimodal sensor fusion according to claim 5, characterized in that, The system includes a multi-mechanism coordination mechanism to automatically fuse and adjust results when short-term, periodic, and long-term adaptive parameters conflict. Specifically, this includes: Weighted fusion formula: Final threshold = w1 × short-term threshold + w2 × periodic threshold + w3 × long-term threshold Where w1 + w2 + w3 = 1, and each weight is automatically assigned according to the recovery stage; Weight dynamic adjustment rules: When the system detects that a user's gait is unstable or that they are fatigued, it increases the weight of the short-term threshold to ensure safety. When a trend of improvement is detected, increase the long-term weight to promote recovery goals; When conflicts among multiple mechanisms cannot be reconciled, the lower bound of safety should be the priority in ensuring safety.

Citation Information

Patent Citations

  • Gait monitoring and muscle force monitoring fusion algorithm and rehabilitation walking aid using same

    CN120899233A

Cited By

  • Method and system for analyzing pumpability of geopolymer concrete

    CN122329920A

  • A method and system for analyzing the pumpability of geopolymer concrete

    CN122329920B