Training risk prediction system based on multi-modal data fusion

Abstract

The application discloses a training risk prediction system based on multi-modal data fusion, comprising: a multi-source signal acquisition module synchronously acquires heart rate, three-axis acceleration and body surface temperature time series signals, and obtains a multi-modal original signal sequence through clock synchronization and abnormality repair; a characteristic vector generation module completes signal filtering, window segmentation and characteristic standardization, and outputs a multi-modal characteristic vector set; an individualized baseline modeling module relies on resting homologous signals, and constructs an exclusive physiological baseline model through multi-element Gaussian distribution fitting; a risk deviation degree calculation module solves static deviation degree by using Mahalanobis distance, and generates a window risk comprehensive score by fusing dynamic indexes; and a risk trend research and judgment module smoothes risk time series data, matches multi-level thresholds, and outputs fatigue accumulation and sudden damage early warning. The application builds an individualized physiological judgment benchmark, quantifies training risk by fusing dynamic and static indexes, can distinguish progressive fatigue and instantaneous damage hidden dangers, and effectively reduces risk misjudgment and missed judgment.

Description

Technical Field

[0001] This invention relates to the field of training risk prediction technology, specifically a training risk prediction system based on multimodal data fusion. Background Technology

[0002] In various high-intensity training scenarios, such as physical fitness training, professional sports training, and military training, trainees are highly susceptible to safety risks such as physical exhaustion, accumulated fatigue, and sports injuries. Sudden training injuries not only interrupt the training process but also cause irreversible damage to the trainee's health, and in severe cases, may even lead to safety accidents. Therefore, real-time monitoring and risk prediction of the human physiological state during training are crucial for ensuring training safety and regulating training intensity.

[0003] Traditional training risk monitoring methods are mainly divided into two categories: manual subjective judgment and single physiological signal monitoring. Among them, the manual judgment method relies on the coach's experience and subjectively judges the fatigue level and risk level by observing the trainee's posture and condition. It is highly subjective, has low judgment accuracy, cannot quantify the degree of physiological fatigue accumulation, and is difficult to predict potential injury risks in advance. The single physiological signal monitoring method mostly collects only the heart rate signal for risk assessment and simply determines the training load by the heart rate threshold. It has a single monitoring dimension, poor anti-interference ability, and cannot comprehensively reflect the multidimensional physiological state of the human body during exercise, such as heart rate changes, limb movement amplitude, and body surface temperature changes.

[0004] Existing technical solutions have many inherent flaws: First, the synchronous acquisition and clock alignment of multimodal physiological signals were not achieved, and the timing of heart rate, movement posture and body temperature signal acquisition was misaligned, which could easily lead to distortion of feature analysis. Secondly, there is a lack of individualized physiological baseline modeling mechanisms. The general approach of using a fixed risk threshold to determine the risk of all training subjects is not suitable for the individual physiological differences of different groups. As a result, the risk assessment has poor universality and a high misjudgment rate. Third, the risk assessment model is too simplistic, relying solely on static physiological thresholds to determine risk without combining dynamic motion indicators and physiological fluctuation characteristics, thus failing to distinguish between the risk of chronic fatigue accumulation and the risk of sudden, instantaneous injury. Fourth, the risk assessment lacks the ability to analyze time-series trends, and only makes instantaneous judgments based on single-frame data. It cannot track the fatigue accumulation process, making it difficult to achieve early warning. It can only provide post-event risk alerts, resulting in poor timeliness of the warnings.

[0005] In summary, existing training risk monitoring technologies suffer from problems such as limited monitoring dimensions, lack of individual adaptability, low accuracy of risk assessment, inability to distinguish risk types, and delayed early warning, making it difficult to meet the needs for refined, individualized, and real-time risk prediction and early warning in high-intensity, routine training scenarios.

[0006] It should be noted that the above content falls within the inventor's technical knowledge and does not necessarily constitute prior art. Summary of the Invention

[0007] To overcome the aforementioned deficiencies of existing technologies, this invention provides a training risk prediction system based on multimodal data fusion. Through a complete technical chain of synchronous acquisition of multimodal physiological signals, individualized baseline modeling, dynamic and static risk quantification assessment, and time-series trend filtering analysis, it achieves accurate and real-time prediction and early warning of fatigue accumulation risk and sudden injury risk during training. This solves the technical problems of existing technologies, such as single monitoring dimensions, poor individual adaptability, inaccurate risk judgment, and delayed early warning.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a training risk prediction system based on multimodal data fusion, comprising: Multi-source signal acquisition module: used to synchronously acquire time-series signals of heart rate, triaxial acceleration, and body surface temperature of the training subjects during the training process, and generate time-aligned multimodal raw signal sequences after clock synchronization alignment and outlier repair processing; Feature vector generation module: used to preprocess and window the multimodal raw signal sequence, extract multiple physiological features and standardize them to generate a standardized multimodal feature vector set; Individualized baseline modeling module: used to collect baseline multimodal signals of training subjects in the resting state, and construct individualized physiological response baseline models after homologous processing; Risk deviation calculation module: used to calculate static risk deviation by combining the baseline model and matching dynamic auxiliary indicators to obtain a window risk comprehensive score; Risk trend analysis and early warning output module: It is used to generate a risk trend sequence based on the time series risk comprehensive score, and after filtering, it compares with multi-level thresholds to output the corresponding fatigue accumulation or sudden damage risk early warning signal.

[0009] In some embodiments, the process by which the multi-source signal acquisition module generates a time-aligned multimodal raw signal sequence includes: Simultaneously collect heart rate time-series signal, triaxial acceleration time-series signal, and body surface temperature time-series signal of the training subject throughout the entire exercise training process, and control the three types of signals to use a consistent sampling frequency; A global clock synchronization mechanism is used to calibrate and align the timestamps of the three types of timing signals to eliminate timing deviations in the acquisition of multi-source signals. The outlier detection algorithm is used to identify and replace abrupt abnormal data in the signal. The nearest neighbor interpolation method is used to repair locally missing signals, thus completing the outlier repair and finally generating a time-aligned and data-complete multimodal original signal sequence.

[0010] In some embodiments, the feature vector generation module completes the preprocessing, window segmentation, and feature extraction process as follows: Wavelet filtering algorithm is used to filter and denoise the time-aligned multimodal original signal sequence to remove high-frequency motion noise and environmental interference noise. At the same time, an adaptive baseline correction algorithm is used to eliminate the baseline drift problem of physiological signals. The preprocessed continuous signal sequence was segmented using a fixed-duration sliding window segmentation method. The window duration was set to 2-5 seconds and the window overlap rate was 30%-50%, resulting in several equal-length analysis windows. For each analysis window, three core features are extracted: time-domain index of heart rate variability, root mean square value of resultant acceleration amplitude, and mean temperature change rate. The extracted features from all windows are normalized and standardized to unify the feature dimensions and numerical ranges, and then combined to construct a standardized multimodal feature vector set.

[0011] In some embodiments, the time-domain indicators of heart rate variability include the mean cardiac cycle, the standard deviation of cardiac cycle, and the root mean square of the difference between consecutive cardiac cycles, and the combined acceleration amplitude is obtained by real-time synthesis of triaxial acceleration time-series signals.

[0012] In some embodiments, the process by which the individualized baseline modeling module constructs an individualized physiological response baseline model includes: Collect baseline heart rate, baseline triaxial acceleration, and baseline body surface temperature multimodal signals from training subjects under resting, sitting, and no-exercise conditions. The acquisition time should be no less than 30 seconds to ensure the stability of the baseline signal data. Following the same filtering and noise reduction, baseline drift correction, sliding window segmentation, feature extraction, and standardization process as the feature vector generation module, the benchmark multimodal signal is processed to obtain the benchmark multimodal feature vector set. Based on the benchmark multimodal feature vector set, the benchmark mean vector and covariance matrix are obtained by fitting with a multivariate Gaussian distribution algorithm. The mean vector represents the normal physiological baseline level of the training object, and the covariance matrix represents the correlation fluctuation characteristics of each physiological feature. By combining the mean vector and covariance matrix obtained from the fitting, an individualized physiological response baseline model adapted to the individual physiological differences of the training subjects is constructed.

[0013] In some embodiments, the process by which the risk deviation calculation module calculates the window risk comprehensive score includes: The standardized multimodal feature vector set is used to input the feature vector corresponding to each analysis window into the individualized physiological response baseline model one by one. The deviation of each window feature from the baseline model is calculated by the Mahalanobis distance formula to obtain the static risk deviation index of each window. Three dynamic auxiliary indicators corresponding to each analysis window are extracted simultaneously: the standard deviation of the window's internal rate variability, the peak value of the combined acceleration, and the cumulative temperature change. An extreme value normalization algorithm is used to normalize the static risk deviation index and the three dynamic auxiliary indicators respectively, mapping all index values ​​to a unified range of 0-1. The four normalized indicators are summed with equal weights to obtain the comprehensive risk score for the corresponding analysis window.

[0014] In some embodiments, the working process of the risk trend assessment and early warning output module includes: The risk comprehensive scores corresponding to multiple analysis windows in a continuous time series are arranged in chronological order to construct a complete real-time risk trend sequence; A five-point moving average filtering algorithm is used to smooth the risk trend sequence, eliminate the interference of single-window numerical abrupt changes, and optimize the stability of the risk trend curve. Four risk thresholds—low, medium, high, and extremely high—are preset, and the filtered real-time risk trend values ​​are compared with each threshold level one by one. When the risk value remains in the medium-high threshold range and shows an upward trend, a fatigue accumulation risk warning signal is output. When the risk value instantly exceeds an extremely high threshold, a sudden damage risk warning signal is output.

[0015] In some embodiments, the criteria for determining the fatigue accumulation risk warning are: Five or more consecutive windows have a risk score in the medium-risk threshold range, or three or more consecutive windows have a score in the high-risk threshold range. The judgment condition for the sudden damage risk warning is: the risk comprehensive score of any single window instantly reaches the extremely high risk threshold.

[0016] In some embodiments, the sliding window segmentation method adopts a fixed step size sliding cutting method, with a window duration of 3 seconds and a window overlap rate of 40%, ensuring that the data in adjacent analysis windows are continuous, without gaps, and without redundancy or missing data.

[0017] In some embodiments, the Mahalanobis distance calculation step includes: The mean vector and covariance matrix of the individualized physiological response baseline model are retrieved, and the Mahalanobis distance between the real-time window feature vector and the baseline mean vector is calculated as a static risk deviation index.

[0018] Compared with existing technologies, this invention provides a training risk prediction system based on multimodal data fusion, which has the following beneficial effects: 1. This training risk prediction system based on multimodal data fusion constructs a complete closed-loop technical link of multimodal signal acquisition, preprocessing feature extraction, individualized baseline modeling, dynamic and static fusion risk scoring, and time-series trend early warning. It simultaneously integrates three core physiological signals: heart rate, exercise posture, and body surface temperature, breaking through the limitations of traditional single signal monitoring. It comprehensively covers multiple dimensions such as cardiovascular load, limb exercise intensity, and changes in body metabolic temperature during training, greatly improving the comprehensiveness and accuracy of training status perception. 2. This training risk prediction system based on multimodal data fusion introduces an individualized physiological baseline modeling mechanism, abandons the traditional general fixed threshold judgment mode, and fits a unique multivariate Gaussian distribution baseline model based on the resting physiological data of the training subjects. It accurately adapts to the individual physiological differences of training subjects of different ages, physical fitness, and physical condition, thereby reducing the probability of risk misjudgment and omission from the root and greatly improving the individualized adaptability and universality of risk prediction. 3. This training risk prediction system based on multimodal data fusion adopts a quantitative scoring mechanism that integrates static baseline deviation and dynamic physiological auxiliary indicators. It accurately represents the static deviation of physiological state from normal baseline through Mahalanobis distance, and captures instantaneous exercise risk in real time by combining three types of dynamic indicators: heart rate fluctuation, exercise peak, and temperature cumulative change. It achieves dual quantitative assessment of static fatigue accumulation and dynamic sudden risk, solving the problems of traditional technology being unable to distinguish risk types and having a single assessment dimension. 4. This training risk prediction system based on multimodal data fusion introduces time-series moving average filtering and multi-level time-series threshold judgment mechanism to eliminate misjudgment interference caused by single-window data mutations. It identifies the fatigue accumulation process through continuous window time-series trend analysis and judges the risk of sudden damage by instantaneous threshold breakthrough. It achieves a hierarchical early warning effect of early prediction of fatigue accumulation and immediate warning of sudden damage, which solves the defects of traditional technology in early warning lag and lack of trend judgment capability, and effectively ensures training safety. Meanwhile, the entire process of data collection, processing, and analysis is automated, eliminating the need for manual intervention and subjective judgment. The entire process of data preprocessing, feature extraction, model fitting, risk calculation, and early warning output is standardized and digitized, making it suitable for various normalized, high-intensity, and large-scale training scenarios. It has strong scenario adaptability and practical applicability. Attached Figure Description

[0019] Figure 1 This is a system architecture block diagram of the training risk prediction system based on multimodal data fusion according to the present invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments and accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] It should be understood that the step numbers used in the text are for ease of description only and are not intended to limit the order in which the steps are performed.

[0022] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0023] The terms “comprising” and “including” indicate the presence of the described feature, whole, step, operation, element and / or component, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or collections thereof.

[0024] The term “and / or” refers to any combination of one or more of the associated listed items, as well as all possible combinations, and includes these combinations.

[0025] Please see Figure 1 In this implementation plan, the training risk prediction system based on multimodal data fusion includes: a multi-source signal acquisition module, a feature vector generation module, an individualized baseline modeling module, a risk deviation calculation module, and a risk trend judgment and early warning output module.

[0026] Multi-source signal acquisition module: used to synchronously acquire time-series signals of heart rate, triaxial acceleration, and body surface temperature of the training subjects during the training process, and generate time-aligned multimodal raw signal sequences after clock synchronization alignment and outlier repair processing.

[0027] Specifically, the heart rate time-series signal, triaxial acceleration time-series signal, and body surface temperature time-series signal of the training subjects are collected synchronously throughout the exercise training process, and the sampling frequency is uniformly set to 100Hz. Wearable devices use the main control chip clock as a reference to complete the calibration and alignment of timestamps from multiple sensors, eliminating timing deviations caused by local clock drift of different sensors; Box plot criteria are used to identify abrupt and abnormal sampling points within the signal, and abnormal data that exceed the reasonable value range are replaced with boundary values; To address data gaps caused by communication interruptions, a linear interpolation method using adjacent valid sampling points is employed to fill in the missing segments; after anomaly repair, a multimodal original signal sequence with time alignment and continuous data is output.

[0028] Feature vector generation module: used to preprocess and window segment the original multimodal signal sequence, extract and standardize multiple physiological features, and generate a standardized multimodal feature vector set.

[0029] Specifically, a wavelet filter is constructed using the db3 wavelet basis to perform noise reduction on the original signal, filtering out high-frequency artifacts and environmental interference caused by motion; simultaneously, an adaptive baseline correction method is used to eliminate the slow baseline drift phenomenon of physiological signals. A sliding window mechanism is used to segment and analyze continuous signals. In this embodiment, the preferred window duration is 3s and the window overlap rate is 40%. The window sliding step size is calculated based on the above parameters, and overlapping continuous analysis windows are obtained in sequence to ensure that there is no loss of time series features. For each analysis window, multiple key features are extracted simultaneously: a resultant acceleration sequence is synthesized based on triaxial acceleration data, and the root mean square of the resultant acceleration within the window is calculated; multiple time-domain indices of heart rate variability corresponding to the heart rate signal are extracted; and the average rate of change of body surface temperature within the window is statistically analyzed. Extreme value normalization is performed on all extracted features to eliminate differences in the dimensions of different features, mapping all features to a unified numerical range and combining them to form a standardized multimodal feature vector set.

[0030] Individualized baseline modeling module: used to collect baseline multimodal signals of training subjects in the resting state, and construct an individualized physiological response baseline model after homologous processing.

[0031] Specifically, before the formal training begins, baseline physiological signals of the training subjects are collected under conditions of sitting and resting without exercise load, with each collection lasting no less than 30 seconds. The baseline signals strictly reuse the filtering, window segmentation, feature extraction, and normalization processing procedures that are completely consistent with the feature vector generation module to ensure that the baseline features and real-time training features are calculated using the same logic, and to avoid introducing system errors due to inconsistent processing procedures. Based on the baseline feature vector set obtained under resting conditions, a multivariate Gaussian distribution is used for fitting to obtain the baseline mean vector. With the characteristic covariance matrix The mean vector represents the normal physiological baseline level of the subject, and the covariance matrix describes the natural fluctuation correlation between various physiological characteristics. Its multivariate Gaussian distribution probability density formula is as follows: ; in, For feature dimension, The determinant of the covariance matrix is ​​obtained by fitting the matrix. The data is stored and solidified to form a baseline model that adapts to the individual's physiological characteristics.

[0032] Risk Deviation Calculation Module: Used to calculate static risk deviation by combining the baseline model and matching dynamic auxiliary indicators to obtain a comprehensive risk score for the window.

[0033] Specifically, the mean vector and covariance matrix stored in the baseline model are called, and the deviation of the real-time feature vector from the resting baseline is quantified by Mahalanobis distance. Mahalanobis distance is the core evaluation index of this scheme, and the calculation formula is as follows: ; This is a static risk deviation index. Mahalanobis distance can eliminate the assessment bias caused by the correlation between features and the dimension. Compared with Euclidean distance, it is more suitable for judging the degree of deviation of multidimensional physiological state.

[0034] Simultaneously extract three dynamic auxiliary indicators within the window: standard deviation of heart rate variability, peak combined acceleration, and cumulative temperature change; After normalizing the Mahalanobis distance index and the three dynamic indicators, a weighted fusion is performed to obtain the single-window comprehensive risk score, the formula of which is as follows: ; This embodiment uses an equal-weight configuration. ; These are the normalized values ​​of three dynamic auxiliary indicators, and the comprehensive risk score. The larger the value, the higher the current physiological safety risk.

[0035] Risk trend analysis and early warning output module: It is used to generate a risk trend sequence based on the time series risk comprehensive score, and after filtering, it compares with multi-level thresholds to output the corresponding fatigue accumulation or sudden damage risk early warning signal.

[0036] Specifically, the risk comprehensive scores of multiple consecutive windows in the time series are arranged sequentially to construct a risk trend sequence. A five-point moving average filter is used to suppress single-point numerical disturbances. The smoothing filter formula is as follows: ; The sequence's beginning and end boundaries are filled with nearby values ​​to effectively suppress false fluctuations caused by sensor noise and instantaneous motion interference. The system presets four risk thresholds: low, medium, high, and extremely high. It compares the filtered risk values ​​with each threshold to distinguish between two risk modes: **Fatigue Accumulation Risk Warning Criteria:** If the overall risk score for five or more consecutive windows is in the medium-risk range, or three or more consecutive windows are in the high-risk range, fatigue is considered to be accumulating, and a warning signal is output, suggesting a reduction in training load and scheduled rest intervals. **Sudden Injury Risk Warning Criteria:** If the filtered risk score for any single window reaches the extremely high-risk threshold, instantly exceeding the safety boundary, a sudden injury warning signal is immediately output, prompting immediate cessation of high-intensity training and a check of physical condition.

[0037] It should be noted that the window duration, overlap rate, filter parameters, risk thresholds at each level, and indicator weight coefficients in the solution can all be adaptively adjusted by those skilled in the art according to the type of training population (ordinary people / professional athletes) and the intensity of training subjects, and can be flexibly adapted to different application scenarios.

[0038] In the application, several formulas are calculated by removing dimensions and taking their numerical values. The formulas are established by collecting a large amount of data and simulating the most recent real situation. Some coefficients or weights in the formulas are set by those skilled in the art according to the actual situation, so they will not be elaborated here.

[0039] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0040] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A training risk prediction system based on multimodal data fusion, characterized in that, include: Multi-source signal acquisition module: used to synchronously acquire time-series signals of heart rate, triaxial acceleration, and body surface temperature of the training subjects during the training process, and generate time-aligned multimodal raw signal sequences after clock synchronization alignment and outlier repair processing; Feature vector generation module: used to preprocess and window the multimodal raw signal sequence, extract multiple physiological features and standardize them to generate a standardized multimodal feature vector set; Individualized baseline modeling module: used to collect baseline multimodal signals of training subjects in the resting state, and construct individualized physiological response baseline models after homologous processing; Risk deviation calculation module: used to calculate static risk deviation by combining the baseline model and matching dynamic auxiliary indicators to obtain a window risk comprehensive score; Risk trend analysis and early warning output module: It is used to generate a risk trend sequence based on the time series risk comprehensive score, and after filtering, it compares with multi-level thresholds to output the corresponding fatigue accumulation or sudden damage risk early warning signal.

2. The training risk prediction system based on multimodal data fusion according to claim 1, characterized in that, The process by which the multi-source signal acquisition module generates the time-aligned multimodal raw signal sequence includes: Simultaneously collect heart rate time-series signal, triaxial acceleration time-series signal, and body surface temperature time-series signal of the training subject throughout the entire exercise training process, and control the three types of signals to use a consistent sampling frequency; A global clock synchronization mechanism is used to calibrate and align the timestamps of the three types of timing signals to eliminate timing deviations in the acquisition of multi-source signals. The outlier detection algorithm is used to identify and replace abrupt abnormal data in the signal. The nearest neighbor interpolation method is used to repair locally missing signals, thus completing the outlier repair and finally generating a time-aligned and data-complete multimodal original signal sequence.

3. The training risk prediction system based on multimodal data fusion according to claim 1, characterized in that, The feature vector generation module completes the preprocessing, window segmentation, and feature extraction processes, including: Wavelet filtering algorithm is used to filter and denoise the time-aligned multimodal original signal sequence to remove high-frequency motion noise and environmental interference noise. At the same time, an adaptive baseline correction algorithm is used to eliminate the baseline drift problem of physiological signals. The preprocessed continuous signal sequence was segmented using a fixed-duration sliding window segmentation method. The window duration was set to 2-5 seconds and the window overlap rate was 30%-50%, resulting in several equal-length analysis windows. For each analysis window, three core features are extracted: time-domain index of heart rate variability, root mean square value of resultant acceleration amplitude, and mean temperature change rate. The extracted features from all windows are normalized and standardized to unify the feature dimensions and numerical ranges, and then combined to construct a standardized multimodal feature vector set.

4. The training risk prediction system based on multimodal data fusion according to claim 3, characterized in that, The time-domain indicators of heart rate variability include the mean cardiac cycle, the standard deviation of cardiac cycle, and the root mean square of the difference between consecutive cardiac cycles. The combined acceleration amplitude is obtained by real-time synthesis of triaxial acceleration time-series signals.

5. The training risk prediction system based on multimodal data fusion according to claim 1, characterized in that, The process of constructing a personalized physiological response baseline model by the personalized baseline modeling module includes: Collect baseline heart rate, baseline triaxial acceleration, and baseline body surface temperature multimodal signals from training subjects under resting, sitting, and no-exercise conditions. The acquisition time should be no less than 30 seconds to ensure the stability of the baseline signal data. Following the same filtering and noise reduction, baseline drift correction, sliding window segmentation, feature extraction, and standardization process as the feature vector generation module, the benchmark multimodal signal is processed to obtain the benchmark multimodal feature vector set. Based on the benchmark multimodal feature vector set, the benchmark mean vector and covariance matrix are obtained by fitting with a multivariate Gaussian distribution algorithm. The mean vector represents the normal physiological baseline level of the training object, and the covariance matrix represents the correlation fluctuation characteristics of each physiological feature. By combining the mean vector and covariance matrix obtained from the fitting, an individualized physiological response baseline model adapted to the individual physiological differences of the training subjects is constructed.

6. The training risk prediction system based on multimodal data fusion according to claim 1, characterized in that, The process by which the risk deviation calculation module calculates the window risk comprehensive score includes: The standardized multimodal feature vector set is used to input the feature vector corresponding to each analysis window into the individualized physiological response baseline model one by one. The deviation of each window feature from the baseline model is calculated by the Mahalanobis distance formula to obtain the static risk deviation index of each window. Three dynamic auxiliary indicators corresponding to each analysis window are extracted simultaneously: the standard deviation of the window's internal rate variability, the peak value of the combined acceleration, and the cumulative temperature change. An extreme value normalization algorithm is used to normalize the static risk deviation index and the three dynamic auxiliary indicators respectively, mapping all index values ​​to a unified range of 0-1. The four normalized indicators are summed with equal weights to obtain the comprehensive risk score for the corresponding analysis window.

7. The training risk prediction system based on multimodal data fusion according to claim 1, characterized in that, The working process of the risk trend assessment and early warning output module includes: The risk comprehensive scores corresponding to multiple analysis windows in a continuous time series are arranged in chronological order to construct a complete real-time risk trend sequence; A five-point moving average filtering algorithm is used to smooth the risk trend sequence, eliminate the interference of single-window numerical abrupt changes, and optimize the stability of the risk trend curve. Four risk thresholds—low, medium, high, and extremely high—are preset, and the filtered real-time risk trend values ​​are compared with each threshold level one by one. When the risk value remains in the medium-high threshold range and shows an upward trend, a fatigue accumulation risk warning signal is output. When the risk value instantly exceeds an extremely high threshold, a sudden damage risk warning signal is output.

8. The training risk prediction system based on multimodal data fusion according to claim 7, characterized in that, The criteria for determining the fatigue accumulation risk warning are as follows: Five or more consecutive windows have a risk score in the medium-risk threshold range, or three or more consecutive windows have a score in the high-risk threshold range. The judgment condition for the sudden damage risk warning is: the risk comprehensive score of any single window instantly reaches the extremely high risk threshold.

9. The training risk prediction system based on multimodal data fusion according to claim 3, characterized in that, The sliding window segmentation method adopts a fixed step size sliding cutting method with a window duration of 3 seconds and a window overlap rate of 40%, ensuring that the data in adjacent analysis windows are continuous, without gaps, and without redundancy or missing data.

10. The training risk prediction system based on multimodal data fusion according to claim 6, characterized in that, The Mahalanobis distance calculation steps include: The mean vector and covariance matrix of the individualized physiological response baseline model are retrieved, and the Mahalanobis distance between the real-time window feature vector and the baseline mean vector is calculated as a static risk deviation index.

Images